Route generated chunks over transient events

Wire transient router through node lifecycle
feat: add transient event transport
2026-09-08 19:41:32 -04:00 · 2026-05-03 02:24:36 +01:00 · 2026-05-03 02:19:13 +01:00 · 2026-05-03 02:03:58 +01:00 · 2026-05-03 02:01:08 +01:00 · 2026-05-03 01:58:47 +01:00
347 changed files with 23742 additions and 5139 deletions
@@ -1 +1,8 @@
 use flake
+
+# creates .venv if doesn't exist and loads its environment
+export VIRTUAL_ENV=".venv"
+if ! [ -d "./$VIRTUAL_ENV" ]; then
+  uv venv
+fi
+layout python
@@ -32,7 +32,6 @@ jobs:
      SPARKLE_ED25519_PRIVATE: ${{ secrets.SPARKLE_ED25519_PRIVATE }}
      SPARKLE_S3_BUCKET: ${{ secrets.SPARKLE_S3_BUCKET }}
      SPARKLE_S3_PREFIX: ${{ secrets.SPARKLE_S3_PREFIX }}
-      EXO_BUG_REPORT_PRESIGNED_URL_ENDPOINT: ${{ secrets.EXO_BUG_REPORT_PRESIGNED_URL_ENDPOINT }}
      AWS_REGION: ${{ secrets.AWS_REGION }}
      EXO_BUILD_NUMBER: ${{ github.run_number }}
      EXO_LIBP2P_NAMESPACE: ${{ github.ref_name }}
@@ -159,7 +158,7 @@ jobs:
          fi

      - name: Install Homebrew packages
-        run: brew install just awscli macmon
+        run: brew install just awscli

      - name: Install UV
        uses: astral-sh/setup-uv@v6
@@ -239,10 +238,92 @@ jobs:
          # Export keychain path for other steps
          echo "BUILD_KEYCHAIN_PATH=$KEYCHAIN_PATH" >> $GITHUB_ENV

+      # ============================================================
+      # Pre-flight credential / profile validation
+      # Runs BEFORE the ~16 min build so auth/expiry failures surface in <1 min.
+      # ============================================================
+
+      - name: Validate Apple notarization credentials
+        env:
+          APPLE_NOTARIZATION_USERNAME: ${{ secrets.APPLE_NOTARIZATION_USERNAME }}
+          APPLE_NOTARIZATION_PASSWORD: ${{ secrets.APPLE_NOTARIZATION_PASSWORD }}
+          APPLE_NOTARIZATION_TEAM: ${{ secrets.APPLE_NOTARIZATION_TEAM }}
+        run: |
+          # All-or-nothing: either all three creds are set, or none are.
+          CRED_COUNT=0
+          for v in "$APPLE_NOTARIZATION_USERNAME" "$APPLE_NOTARIZATION_PASSWORD" "$APPLE_NOTARIZATION_TEAM"; do
+            [[ -n "$v" ]] && CRED_COUNT=$((CRED_COUNT + 1))
+          done
+          if [[ "$CRED_COUNT" -eq 0 ]]; then
+            echo "No notarization credentials configured — skipping notarization for this build."
+            exit 0
+          fi
+          if [[ "$CRED_COUNT" -ne 3 ]]; then
+            echo "ERROR: partial notarization credentials set ($CRED_COUNT/3). Aborting before build."
+            exit 1
+          fi
+          # Cheap, ~5s, auth-only call. Fails instantly with a clear message if
+          # the app-specific password is stale, wrong team-id, etc.
+          echo "Verifying Apple notarization credentials via notarytool history..."
+          if ! xcrun notarytool history \
+            --apple-id "$APPLE_NOTARIZATION_USERNAME" \
+            --password "$APPLE_NOTARIZATION_PASSWORD" \
+            --team-id "$APPLE_NOTARIZATION_TEAM" >/dev/null; then
+            echo "ERROR: notarytool rejected the provided credentials. Fix before rerunning."
+            echo "Common causes: app-specific password expired/revoked, wrong team-id,"
+            echo "Apple ID not on the team, or 2FA not configured for this Apple ID."
+            exit 1
+          fi
+          echo "Apple notarization credentials OK."
+
+      - name: Validate provisioning profile expiry
+        run: |
+          PROFILE="$HOME/Library/Developer/Xcode/UserData/Provisioning Profiles/EXO.provisionprofile"
+          if [[ ! -f "$PROFILE" ]]; then
+            echo "ERROR: provisioning profile not found at $PROFILE"
+            exit 1
+          fi
+          EXPIRY=$(security cms -D -i "$PROFILE" | plutil -extract ExpirationDate raw -o - - 2>/dev/null || true)
+          if [[ -z "$EXPIRY" ]]; then
+            echo "WARNING: could not read ExpirationDate from provisioning profile; skipping expiry check."
+            exit 0
+          fi
+          # Try a couple of known plutil date formats. If none parse, skip the check rather
+          # than risk a false-positive "expired" block on a format we didn't anticipate.
+          EXPIRY_EPOCH=""
+          for fmt in "%Y-%m-%dT%H:%M:%SZ" "%Y-%m-%d %H:%M:%S %z" "%Y-%m-%d %H:%M:%S +0000"; do
+            if parsed=$(date -j -f "$fmt" "$EXPIRY" +%s 2>/dev/null); then
+              EXPIRY_EPOCH="$parsed"
+              break
+            fi
+          done
+          if [[ -z "$EXPIRY_EPOCH" ]]; then
+            echo "WARNING: could not parse ExpirationDate '$EXPIRY'; skipping expiry check."
+            exit 0
+          fi
+          NOW_EPOCH=$(date +%s)
+          if [[ "$EXPIRY_EPOCH" -le "$NOW_EPOCH" ]]; then
+            echo "ERROR: provisioning profile expired on $EXPIRY. Regenerate it before rerunning."
+            exit 1
+          fi
+          DAYS_LEFT=$(( (EXPIRY_EPOCH - NOW_EPOCH) / 86400 ))
+          echo "Provisioning profile valid until $EXPIRY ($DAYS_LEFT days remaining)."
+          if [[ "$DAYS_LEFT" -lt 14 ]]; then
+            echo "WARNING: profile expires in under 14 days — regenerate soon."
+          fi
+
      # ============================================================
      # Build the bundle
      # ============================================================

+      - name: Add pinned macmon to PATH
+        run: |
+          MACMON_DIR=$(nix develop --command sh -c 'dirname $(which macmon)')
+          echo "Using macmon from: $MACMON_DIR"
+          echo "$MACMON_DIR" >> $GITHUB_PATH
+          # Remove any Homebrew macmon so PyInstaller can't accidentally pick it up
+          brew uninstall macmon 2>/dev/null || true
+
      - name: Build PyInstaller bundle
        run: uv run pyinstaller packaging/pyinstaller/exo.spec

@@ -265,7 +346,6 @@ jobs:
            EXO_BUILD_COMMIT="$GITHUB_SHA" \
            SPARKLE_FEED_URL="$SPARKLE_FEED_URL" \
            SPARKLE_ED25519_PUBLIC="$SPARKLE_ED25519_PUBLIC" \
-            EXO_BUG_REPORT_PRESIGNED_URL_ENDPOINT="$EXO_BUG_REPORT_PRESIGNED_URL_ENDPOINT" \
            CODE_SIGNING_IDENTITY="$SIGNING_IDENTITY" \
            CODE_SIGN_INJECT_BASE_ENTITLEMENTS=YES
          mkdir -p ../../output
@@ -298,11 +378,41 @@ jobs:
          APPLE_NOTARIZATION_PASSWORD: ${{ secrets.APPLE_NOTARIZATION_PASSWORD }}
          APPLE_NOTARIZATION_TEAM: ${{ secrets.APPLE_NOTARIZATION_TEAM }}
        run: |
+          set -o pipefail
          cd output
          security unlock-keychain -p "$MACOS_CERTIFICATE_PASSWORD" "$BUILD_KEYCHAIN_PATH"
          SIGNING_IDENTITY=$(security find-identity -v -p codesigning "$BUILD_KEYCHAIN_PATH" | awk -F '"' '{print $2}')
+
+          # Fail fast if notarization creds are partial. All-or-nothing.
+          CRED_COUNT=0
+          for v in "$APPLE_NOTARIZATION_USERNAME" "$APPLE_NOTARIZATION_PASSWORD" "$APPLE_NOTARIZATION_TEAM"; do
+            [[ -n "$v" ]] && CRED_COUNT=$((CRED_COUNT + 1))
+          done
+          if [[ "$CRED_COUNT" -ne 0 && "$CRED_COUNT" -ne 3 ]]; then
+            echo "ERROR: partial Apple notarization credentials set ($CRED_COUNT/3). Aborting."
+            exit 1
+          fi
+
          /usr/bin/codesign --deep --force --timestamp --options runtime \
            --sign "$SIGNING_IDENTITY" EXO.app
+
+          # Pre-flight: verify the signed app BEFORE building DMG and submitting to Apple.
+          # If this fails, notarization will fail too — cheap way to fail in seconds, not 15 minutes.
+          echo "===== codesign --verify EXO.app ====="
+          if ! /usr/bin/codesign --verify --deep --strict --verbose=2 EXO.app; then
+            echo "ERROR: EXO.app failed codesign verification. Dumping signing status of every executable:"
+            find EXO.app -type f \( -perm -111 -o -name "*.dylib" -o -name "*.so" -o -name "*.framework" \) -print0 |
+              while IFS= read -r -d '' f; do
+                printf -- '--- %s\n' "$f"
+                /usr/bin/codesign -dv --verbose=2 "$f" 2>&1 | sed 's/^/    /' || true
+              done
+            exit 1
+          fi
+
+          # Gatekeeper assessment. A failure here strongly predicts notarization rejection.
+          echo "===== spctl assessment (predicts notarization outcome) ====="
+          /usr/bin/spctl -a -vvv -t install EXO.app || echo "WARNING: spctl assessment failed — notarization is likely to fail too."
+
          mkdir -p dmg-root
          cp -R EXO.app dmg-root/
          ln -s /Applications dmg-root/Applications
@@ -310,12 +420,22 @@ jobs:
          hdiutil create -volname "EXO" -srcfolder dmg-root -ov -format UDZO "$DMG_NAME"
          /usr/bin/codesign --force --timestamp --options runtime \
            --sign "$SIGNING_IDENTITY" "$DMG_NAME"
+
+          echo "===== codesign --verify DMG ====="
+          if ! /usr/bin/codesign --verify --verbose=2 "$DMG_NAME"; then
+            echo "ERROR: DMG failed codesign verification."
+            exit 1
+          fi
+
          if [[ -n "$APPLE_NOTARIZATION_USERNAME" ]]; then
+            echo "===== notarytool submit ====="
+            # `|| true` so set -e doesn't abort before we can echo output / fetch the log.
+            # We rely on the parsed STATUS below to decide pass/fail.
            SUBMISSION_OUTPUT=$(xcrun notarytool submit "$DMG_NAME" \
              --apple-id "$APPLE_NOTARIZATION_USERNAME" \
              --password "$APPLE_NOTARIZATION_PASSWORD" \
              --team-id "$APPLE_NOTARIZATION_TEAM" \
-              --wait --timeout 15m 2>&1)
+              --wait --timeout 15m 2>&1) || true
            echo "$SUBMISSION_OUTPUT"

            SUBMISSION_ID=$(echo "$SUBMISSION_OUTPUT" | awk 'tolower($1)=="id:" && $2 ~ /^[0-9a-fA-F-]+$/ {print $2; exit}')
@@ -91,9 +91,6 @@ jobs:
            nix build .#metal-toolchain
          fi

-          # Build mlx (depends on metal-toolchain)
-          nix build .#mlx
-
      - name: Build all Nix outputs
        run: |
          nix flake show --json | jq -r '
@@ -38,3 +38,5 @@ bench/**/*.json

 # tmp
 tmp/models
+/build/exo
+/.claude/skills
@@ -1,31 +0,0 @@
-<?xml version="1.0" encoding="UTF-8"?>
-<module type="EMPTY_MODULE" version="4">
-  <component name="FacetManager">
-    <facet type="Python" name="Python facet">
-      <configuration sdkName="Python 3.13 virtualenv at ~/Desktop/exo/.venv" />
-    </facet>
-  </component>
-  <component name="Go" enabled="true" />
-  <component name="NewModuleRootManager">
-    <content url="file://$MODULE_DIR$">
-      <sourceFolder url="file://$MODULE_DIR$/scripts/src" isTestSource="false" />
-      <sourceFolder url="file://$MODULE_DIR$/src" isTestSource="false" />
-      <sourceFolder url="file://$MODULE_DIR$/rust/exo_pyo3_bindings/src" isTestSource="false" />
-      <sourceFolder url="file://$MODULE_DIR$/rust/exo_pyo3_bindings/tests" isTestSource="true" />
-      <sourceFolder url="file://$MODULE_DIR$/rust/util/src" isTestSource="false" />
-      <sourceFolder url="file://$MODULE_DIR$/rust/networking/examples" isTestSource="false" />
-      <sourceFolder url="file://$MODULE_DIR$/rust/networking/src" isTestSource="false" />
-      <sourceFolder url="file://$MODULE_DIR$/rust/networking/tests" isTestSource="true" />
-      <sourceFolder url="file://$MODULE_DIR$/rust/system_custodian/src" isTestSource="false" />
-      <excludeFolder url="file://$MODULE_DIR$/.venv" />
-      <excludeFolder url="file://$MODULE_DIR$/.direnv" />
-      <excludeFolder url="file://$MODULE_DIR$/build" />
-      <excludeFolder url="file://$MODULE_DIR$/dist" />
-      <excludeFolder url="file://$MODULE_DIR$/.go_cache" />
-      <excludeFolder url="file://$MODULE_DIR$/rust/target" />
-    </content>
-    <orderEntry type="jdk" jdkName="Python 3.13 (exo)" jdkType="Python SDK" />
-    <orderEntry type="sourceFolder" forTests="false" />
-    <orderEntry type="library" name="Python 3.13 virtualenv at ~/Desktop/exo/.venv interpreter library" level="application" />
-  </component>
-</module>
@@ -1,6 +1,6 @@
 <?xml version="1.0" encoding="UTF-8"?>
 <project version="4">
  <component name="ExternalDependencies">
-    <plugin id="systems.fehn.intellijdirenv" />
+    <plugin id="al.aoli.intellijdirenv" />
  </component>
 </project>
@@ -1,14 +0,0 @@
-<component name="InspectionProjectProfileManager">
-  <profile version="1.0">
-    <option name="myName" value="Project Default" />
-    <inspection_tool class="PyCompatibilityInspection" enabled="true" level="WARNING" enabled_by_default="true">
-      <option name="ourVersions">
-        <value>
-          <list size="1">
-            <item index="0" class="java.lang.String" itemvalue="3.14" />
-          </list>
-        </value>
-      </option>
-    </inspection_tool>
-  </profile>
-</component>
@@ -4,7 +4,4 @@
    <option name="sdkName" value="Python 3.13 (exo)" />
  </component>
  <component name="ProjectRootManager" version="2" project-jdk-name="Python 3.13 (exo)" project-jdk-type="Python SDK" />
-  <component name="PythonCompatibilityInspectionAdvertiser">
-    <option name="version" value="3" />
-  </component>
 </project>
@@ -1767,12 +1767,12 @@ def clip(
        array: The clipped array.
    """

-def compile(
-    fun: Callable,
+def compile[F: Callable[..., object]](
+    fun: F,
    inputs: object | None = ...,
    outputs: object | None = ...,
    shapeless: bool = ...,
-) -> Callable:
+) -> F:
    """
    Returns a compiled function which produces the same output as ``fun``.

@@ -2915,8 +2915,8 @@ def gather_mm(
    a: array,
    b: array,
    /,
-    lhs_indices: array,
-    rhs_indices: array,
+    lhs_indices: array | None = ...,
+    rhs_indices: array | None = ...,
    *,
    sorted_indices: bool = ...,
    stream: Stream | Device | None = ...,
@@ -4707,6 +4707,7 @@ def softmax(
    /,
    axis: int | Sequence[int] | None = ...,
    *,
+    precise: bool = ...,
    stream: Stream | Device | None = ...,
 ) -> array:
    """
@@ -57,6 +57,10 @@ class Module(dict):
    def __init__(self) -> None:
        """Should be called by the subclasses of ``Module``."""

+    def __getitem__(self, key: str) -> mx.array | Module: ...
+    def get(
+        self, key: str, default: mx.array | Module | None = ...
+    ) -> mx.array | Module | None: ...
    @property
    def training(self):  # -> bool:
        """Boolean indicating if the model is in training mode."""
@@ -2,7 +2,7 @@
 This type stub file was generated by pyright.
 """

-from typing import Callable, Optional, Union
+from typing import Any, Callable, Optional, Union

 import mlx.core as mx
 from base import Module
@@ -13,8 +13,10 @@ def quantize(
    bits: int = ...,
    *,
    mode: str = ...,
-    class_predicate: Optional[Callable[[str, Module], Union[bool, dict]]] = ...,
-):  # -> None:
+    class_predicate: Optional[
+        Callable[[str, Module], Union[bool, dict[str, Any]]]
+    ] = ...,
+) -> None:
    """Quantize the sub-modules of a module according to a predicate.

    By default all layers that define a ``to_quantized(group_size, bits)``
@@ -0,0 +1,5 @@
+"""
+This type stub file was generated by pyright.
+"""
+
+__version__ = ...
@@ -3,13 +3,12 @@ This type stub file was generated by pyright.
 """

 import contextlib
-from dataclasses import dataclass
-from typing import Any, Callable, Generator, List, Optional, Tuple, Union
-
 import mlx.core as mx
 import mlx.nn as nn
+from dataclasses import dataclass
+from collections import deque
+from typing import Any, Callable, Generator, List, Optional, Sequence, Tuple, Union
 from transformers import PreTrainedTokenizer
-
 from .tokenizer_utils import TokenizerWrapper

 DEFAULT_PROMPT = ...
@@ -29,6 +28,7 @@ def str2bool(string):  # -> bool:
    ...
 def setup_arg_parser():  # -> ArgumentParser:
    """Set up and return the argument parser."""
+    ...

 generation_stream: mx.Stream

@@ -43,6 +43,7 @@ def wired_limit(
    async eval could be running pass in the streams to synchronize with prior
    to exiting the context manager.
    """
+    ...
@dataclass
 class GenerationResponse:
    """
@@ -91,7 +92,7 @@ def generate_step(
    kv_bits: Optional[int] = ...,
    kv_group_size: int = ...,
    quantized_kv_start: int = ...,
-    prompt_progress_callback: Optional[Callable[[int], int]] = ...,
+    prompt_progress_callback: Optional[Callable[[int, int], None]] = ...,
    input_embeddings: Optional[mx.array] = ...,
 ) -> Generator[Tuple[mx.array, mx.array], None, None]:
    """
@@ -117,7 +118,7 @@ def generate_step(
        kv_group_size (int): Group size for KV cache quantization. Default: ``64``.
        quantized_kv_start (int): Step to begin using a quantized KV cache.
           when ``kv_bits`` is non-None. Default: ``0``.
-        prompt_progress_callback (Callable[[int], int]): A call-back which takes the
+        prompt_progress_callback (Callable[[int, int], None]): A call-back which takes the
           prompt tokens processed so far and the total number of prompt tokens.
        input_embeddings (mx.array, optional): Input embeddings to use instead of or in
          conjunction with prompt tokens. Default: ``None``.
@@ -125,6 +126,7 @@ def generate_step(
    Yields:
        Tuple[mx.array, mx.array]: One token and a vector of log probabilities.
    """
+    ...

 def speculative_generate_step(
    prompt: mx.array,
@@ -170,6 +172,7 @@ def speculative_generate_step(
        Tuple[mx.array, mx.array, bool]: One token, a vector of log probabilities,
          and a bool indicating if the token was generated by the draft model
    """
+    ...

 def stream_generate(
    model: nn.Module,
@@ -177,7 +180,7 @@ def stream_generate(
    prompt: Union[str, mx.array, List[int]],
    max_tokens: int = ...,
    draft_model: Optional[nn.Module] = ...,
-    **kwargs: object,
+    **kwargs: Any,
 ) -> Generator[GenerationResponse, None, None]:
    """
    A generator producing text based on the given prompt from the model.
@@ -199,6 +202,7 @@ def stream_generate(
        GenerationResponse: An instance containing the generated text segment and
            associated metadata. See :class:`GenerationResponse` for details.
    """
+    ...

 def generate(
    model: nn.Module,
@@ -219,6 +223,9 @@ def generate(
       kwargs: The remaining options get passed to :func:`stream_generate`.
          See :func:`stream_generate` for more details.
    """
+    ...
+
+def _merge_caches(caches: List[List[Any]]) -> List[Any]: ...
@dataclass
 class BatchStats:
    """
@@ -242,10 +249,263 @@ class BatchStats:
    generation_time: float = ...
    peak_memory: float = ...

+class SequenceStateMachine:
+    """A state machine that uses one Aho-Corasick trie per state to efficiently
+    track state across a generated sequence.
+
+    The transitions are provided as state -> [(sequence, new_state)].
+
+    Example:
+
+        sm = SequenceStateMachine(
+            transitions={
+                "normal": [
+                    (think_start_tokens, "reasoning"),
+                    (tool_start_tokens, "tool"),
+                    (eos, None),
+                ],
+                "reasoning": [
+                    (think_end_tokens, "normal"),
+                    (eos, None),
+                ],
+                "tool": [
+                    (tool_end_tokens, None),
+                    (eos, None)
+                ],
+            },
+            initial="normal"
+        )
+    """
+    def __init__(self, transitions=..., initial=...) -> None: ...
+    def __deepcopy__(self, memo):  # -> SequenceStateMachine:
+        ...
+    def make_state(self):  # -> tuple[str, Any, dict[Any, Any]]:
+        ...
+    @staticmethod
+    def match(state, x):  # -> tuple[tuple[Any, Any | None, Any], Any | None, Any]:
+        ...
+
+class PromptProcessingBatch:
+    """
+    A batch processor for prompt tokens with support for incremental processing.
+
+    This class handles batched prompt processing, managing KV caches and preparing
+    tokens for generation. It supports extending, filtering, and splitting batches.
+    """
+    @dataclass
+    class Response:
+        uid: int
+        progress: tuple
+        end_of_segment: bool
+        end_of_prompt: bool
+        ...
+
+    def __init__(
+        self,
+        model: nn.Module,
+        uids: List[int],
+        caches: List[List[Any]],
+        tokens: Optional[List[List[int]]] = ...,
+        prefill_step_size: int = ...,
+        samplers: Optional[List[Callable[[mx.array], mx.array]]] = ...,
+        fallback_sampler: Optional[Callable[[mx.array], mx.array]] = ...,
+        logits_processors: Optional[
+            List[List[Callable[[mx.array, mx.array], mx.array]]]
+        ] = ...,
+        state_machines: Optional[List[SequenceStateMachine]] = ...,
+        max_tokens: Optional[List[int]] = ...,
+    ) -> None: ...
+    def __len__(self):  # -> int:
+        ...
+    def extract_cache(self, idx: int) -> List[Any]: ...
+    def extend(self, batch):  # -> None:
+        ...
+    def split(self, indices: List[int]):  # -> Self:
+        ...
+    def filter(self, keep: List[int]):  # -> None:
+        ...
+    def prompt(self, tokens: List[List[int]]):  # -> None:
+        """
+        Process prompt tokens through the model.
+
+        Args:
+            tokens: List of token sequences to process.
+        """
+        ...
+
+    def generate(self, tokens: List[List[int]]):  # -> GenerationBatch:
+        """
+        Transition from prompt processing to generation.
+
+        Args:
+            tokens: Final tokens for each sequence to start generation.
+
+        Returns:
+            A GenerationBatch ready for token generation.
+        """
+        ...
+
+    @classmethod
+    def empty(
+        cls,
+        model: nn.Module,
+        fallback_sampler: Callable[[mx.array], mx.array],
+        prefill_step_size: int = ...,
+    ):  # -> Self:
+        ...
+
+class GenerationBatch:
+    """
+    A batched token generator that manages multiple sequences in parallel.
+
+    This class handles the generation phase after prompt processing, managing
+    KV caches, sampling, and stop sequence detection for multiple sequences.
+    """
+    @dataclass
+    class Response:
+        uid: int
+        token: int
+        logprobs: mx.array
+        finish_reason: Optional[str]
+        current_state: Optional[str]
+        match_sequence: Optional[List[int]]
+        prompt_cache: Optional[List[Any]]
+        all_tokens: Optional[List[int]]
+        ...
+
+    model: nn.Module
+    uids: List[int]
+    prompt_cache: List[Any]
+    tokens: List[List[int]]
+    samplers: Optional[List[Callable[[mx.array], mx.array]]]
+    fallback_sampler: Callable[[mx.array], mx.array]
+    logits_processors: Optional[List[List[Callable[[mx.array, mx.array], mx.array]]]]
+    state_machines: List[SequenceStateMachine]
+    max_tokens: List[int]
+    _current_tokens: Optional[mx.array]
+    _current_logprobs: mx.array | List[mx.array]
+    _next_tokens: Optional[mx.array]
+    _next_logprobs: mx.array | List[mx.array]
+    _token_context: List[Any]
+    _num_tokens: List[int]
+    _matcher_states: List[Any]
+
+    def __init__(
+        self,
+        model: nn.Module,
+        uids: List[int],
+        inputs: mx.array,
+        prompt_cache: List[Any],
+        tokens: List[List[int]],
+        samplers: Optional[List[Callable[[mx.array], mx.array]]],
+        fallback_sampler: Callable[[mx.array], mx.array],
+        logits_processors: Optional[
+            List[List[Callable[[mx.array, mx.array], mx.array]]]
+        ],
+        state_machines: List[SequenceStateMachine],
+        max_tokens: List[int],
+    ) -> None: ...
+    def __len__(self) -> int: ...
+    def extend(self, batch: GenerationBatch) -> None: ...
+    def extract_cache(self, idx: int) -> List[Any]: ...
+    def filter(self, keep: List[int]) -> None: ...
+    def _step(self) -> Tuple[List[int], List[mx.array]]: ...
+    def next(self) -> List[Response]:
+        """
+        Generate the next batch of tokens.
+
+        Returns:
+            List of Response objects for each sequence in the batch.
+        """
+        ...
+
+    @classmethod
+    def empty(
+        cls, model: nn.Module, fallback_sampler: Callable[[mx.array], mx.array]
+    ):  # -> Self:
+        ...
+
+class BatchGenerator:
+    """
+    A batch generator implements continuous batching.
+
+    This class provides automatic management of prompt processing and generation
+    batches, handling the transition between the two.
+
+    It also allows for segmented prompt processing which guarantees that the
+    generator will stop at these boundaries when processing an input.
+    """
+    def __init__(
+        self,
+        model: nn.Module,
+        max_tokens: int = ...,
+        stop_tokens: Optional[Sequence[Sequence[int]]] = ...,
+        sampler: Optional[Callable[[mx.array], mx.array]] = ...,
+        logits_processors: Optional[
+            List[Callable[[mx.array, mx.array], mx.array]]
+        ] = ...,
+        completion_batch_size: int = ...,
+        prefill_batch_size: int = ...,
+        prefill_step_size: int = ...,
+    ) -> None: ...
+    def close(self) -> None: ...
+    def __del__(self):  # -> None:
+        ...
+    @contextlib.contextmanager
+    def stats(self, stats=...):  # -> Generator[Any | BatchStats, Any, None]:
+        ...
+    _unprocessed_sequences: deque[tuple[Any, ...]]
+    _prompt_batch: PromptProcessingBatch
+    _generation_batch: GenerationBatch
+    _currently_processing: list[Any]
+    _gen_tokens_counter: int
+    _steps_counter: int
+    def _next(
+        self,
+    ) -> tuple[
+        List[PromptProcessingBatch.Response], List[GenerationBatch.Response]
+    ]: ...
+    def insert(
+        self,
+        prompts: List[List[int]],
+        max_tokens: Optional[List[int]] = ...,
+        caches: Optional[List[List[Any]]] = ...,
+        all_tokens: Optional[List[List[int]]] = ...,
+        samplers: Optional[List[Callable[[mx.array], mx.array]]] = ...,
+        logits_processors: Optional[
+            List[List[Callable[[mx.array, mx.array], mx.array]]]
+        ] = ...,
+        state_machines: Optional[List[SequenceStateMachine]] = ...,
+    ) -> List[int]: ...
+    def insert_segments(
+        self,
+        segments: List[List[List[int]]],
+        max_tokens: Optional[List[int]] = ...,
+        caches: Optional[List[List[Any]]] = ...,
+        all_tokens: Optional[List[List[int]]] = ...,
+        samplers: Optional[List[Callable[[mx.array], mx.array]]] = ...,
+        logits_processors: Optional[
+            List[List[Callable[[mx.array, mx.array], mx.array]]]
+        ] = ...,
+        state_machines: Optional[List[SequenceStateMachine]] = ...,
+    ) -> List[int]: ...
+    def extract_cache(self, uids: List[int]) -> dict[int, Any]: ...
+    def remove(
+        self, uids: List[int], return_prompt_caches: bool = ...
+    ) -> dict[int, Any]: ...
+    @property
+    def prompt_cache_nbytes(self) -> int: ...
+    def next(
+        self,
+    ) -> tuple[
+        List[PromptProcessingBatch.Response], List[GenerationBatch.Response]
+    ]: ...
+    def next_generated(self) -> List[GenerationBatch.Response]: ...
+
@dataclass
 class BatchResponse:
    """
-    An data object to hold a batch generation response.
+    A data object to hold a batch generation response.

    Args:
        texts: (List[str]): The generated text for each prompt.
@@ -255,98 +515,17 @@ class BatchResponse:
    texts: List[str]
    stats: BatchStats
    caches: Optional[List[List[Any]]]
-
-def _left_pad_prompts(prompts: Any, max_length: Optional[int] = ...) -> mx.array: ...
-def _right_pad_prompts(prompts: Any, max_length: Optional[int] = ...) -> mx.array: ...
-def _make_cache(
-    model: Any, left_padding: Any, max_kv_size: Optional[int]
-) -> List[Any]: ...
-def _merge_caches(caches: Any) -> List[Any]: ...
-@dataclass
-class Batch:
-    uids: List[int]
-    y: mx.array
-    logprobs: List[mx.array] | mx.array
-    max_tokens: List[int]
-    num_tokens: List[int]
-    cache: List[Any]
-    samplers: List[Callable[[mx.array], mx.array] | None]
-    logits_processors: List[List[Callable[[mx.array, mx.array], mx.array]]]
-    tokens: List[mx.array]
-    def __len__(self) -> int: ...
-    def filter(self, keep_idx: List[int]) -> None: ...
-    def extend(self, other: "Batch") -> None: ...
-    def extract_cache(self, idx: int) -> List[Any]: ...
-
-class BatchGenerator:
-    model: nn.Module
-    sampler: Callable[[mx.array], mx.array]
-    stop_tokens: set[int]
-    max_kv_size: Optional[int]
-    prefill_step_size: int
-    completion_batch_size: int
-    prefill_batch_size: int
-    unprocessed_prompts: List[Any]
-    active_batch: Optional[Batch]
-    prompt_progress_callback: Callable[[List[Tuple[int, int, int]]], None]
-    _stats: BatchStats
-    _next_count: int
-
-    @dataclass
-    class Response:
-        uid: int
-        token: int
-        logprobs: mx.array
-        finish_reason: Optional[str]
-        prompt_cache: Any
-
-    def __init__(
-        self,
-        model: Any,
-        max_tokens: int = ...,
-        stop_tokens: Optional[set[int]] = ...,
-        sampler: Optional[Callable[[mx.array], mx.array]] = ...,
-        logits_processors: Optional[
-            List[Callable[[mx.array, mx.array], mx.array]]
-        ] = ...,
-        completion_batch_size: int = ...,
-        prefill_batch_size: int = ...,
-        prefill_step_size: int = ...,
-        prompt_progress_callback: Optional[
-            Callable[[List[Tuple[int, int, int]]], None]
-        ] = ...,
-        max_kv_size: Optional[int] = ...,
-    ) -> None: ...
-    def close(self) -> None: ...
-    def insert(
-        self,
-        prompts: Any,
-        max_tokens: Union[List[int], int, None] = ...,
-        caches: Any = ...,
-        samplers: Optional[List[Any]] = ...,
-        logits_processors: Optional[List[Any]] = ...,
-    ) -> List[int]: ...
-    def remove(
-        self, uids: List[int], return_prompt_caches: bool = ...
-    ) -> Optional[dict[int, List[Any]]]: ...
-    def stats(self) -> BatchStats: ...
-    def next(self) -> List[Response]: ...
-    def _process_prompts(self, prompts: List[Any]) -> Batch: ...
-    def _step(
-        self,
-        input_tokens: mx.array,
-        prompt_cache: List[Any],
-        samplers: Optional[List[Any]],
-        logits_processors: Optional[List[Any]],
-        tokens: List[mx.array],
-    ) -> Tuple[mx.array, List[mx.array]]: ...
+    ...

 def batch_generate(
    model,
    tokenizer,
-    prompts: List[int],
+    prompts: List[List[int]],
+    prompt_caches: Optional[List[List[Any]]] = ...,
    max_tokens: Union[int, List[int]] = ...,
    verbose: bool = ...,
+    return_prompt_caches: bool = ...,
+    logits_processors: Optional[List[Callable[[mx.array, mx.array], mx.array]]] = ...,
    **kwargs,
 ) -> BatchResponse:
    """
@@ -355,14 +534,22 @@ def batch_generate(
    Args:
       model (nn.Module): The language model.
       tokenizer (PreTrainedTokenizer): The tokenizer.
-       prompt (List[List[int]]): The input prompts.
+       prompts (List[List[int]]): The input prompts.
+       prompt_caches (List[List[Any]], optional): Pre-computed prompt-caches
+          for each input prompt. Note, unlike ``generate_step``, the caches
+          won't be updated in-place.
       verbose (bool): If ``True``, print tokens and timing information.
          Default: ``False``.
       max_tokens (Union[int, List[int]): Maximum number of output tokens. This
          can be per prompt if a list is provided.
+       return_prompt_caches (bool): Return the prompt caches in the batch
+          responses. Default: ``False``.
+       logits_processors (List[Callable[[mx.array, mx.array], mx.array]], optional):
+          A list of functions that take tokens and logits and return the processed logits. Default: ``None``.
       kwargs: The remaining options get passed to :obj:`BatchGenerator`.
          See :obj:`BatchGenerator` for more details.
    """
+    ...

 def main():  # -> None:
    ...
@@ -0,0 +1,19 @@
+"""
+This type stub file was generated by pyright.
+"""
+
+import mlx.core as mx
+import mlx.nn as nn
+from functools import partial
+
+@partial(mx.compile, shapeless=True)
+def swiglu(gate, x): ...
+@partial(mx.compile, shapeless=True)
+def xielu(x, alpha_p, alpha_n, beta, eps):  # -> array:
+    ...
+
+class XieLU(nn.Module):
+    def __init__(
+        self, alpha_p_init=..., alpha_n_init=..., beta=..., eps=...
+    ) -> None: ...
+    def __call__(self, x: mx.array) -> mx.array: ...
@@ -3,7 +3,7 @@ This type stub file was generated by pyright.
 """

 from dataclasses import dataclass
-from typing import Optional
+from typing import Any, Optional

 import mlx.core as mx

@@ -37,10 +37,10 @@ def quantized_scaled_dot_product_attention(
    bits: int = ...,
 ) -> mx.array: ...
 def scaled_dot_product_attention(
-    queries,
-    keys,
-    values,
-    cache,
+    queries: mx.array,
+    keys: mx.array,
+    values: mx.array,
+    cache: Optional[Any],
    scale: float,
    mask: Optional[mx.array],
    sinks: Optional[mx.array] = ...,
@@ -0,0 +1,280 @@
+"""Type stubs for mlx_lm.models.deepseek_v4"""
+
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .base import BaseModelArgs
+from .cache import ArraysCache, RotatingKVCache
+from .switch_layers import SwitchGLU
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    vocab_size: int
+    hidden_size: int
+    intermediate_size: int
+    moe_intermediate_size: int
+    num_hidden_layers: int
+    num_attention_heads: int
+    num_key_value_heads: int
+    n_shared_experts: Optional[int]
+    n_routed_experts: int
+    num_experts_per_tok: int
+    head_dim: int
+    qk_rope_head_dim: int
+    q_lora_rank: int
+    o_lora_rank: int
+    o_groups: int
+    sliding_window: int
+    hc_mult: int
+    hc_sinkhorn_iters: int
+    hc_eps: float
+    compress_ratios: Optional[List[int]]
+    compress_rope_theta: float
+    rope_theta: float
+    rope_scaling: Optional[Dict[str, Any]]
+    rms_norm_eps: float
+    swiglu_limit: float
+    attention_bias: bool
+    max_position_embeddings: int
+
+class DeepseekV4RoPE(nn.Module):
+    dims: int
+    freqs: mx.array
+
+    def __init__(
+        self,
+        dims: int,
+        base: float,
+        scaling_config: Optional[Dict[str, Any]] = None,
+    ) -> None: ...
+    def __call__(
+        self,
+        x: mx.array,
+        offset: int = 0,
+        inverse: bool = False,
+    ) -> mx.array: ...
+
+class HyperConnection(nn.Module):
+    dim: int
+    hc_mult: int
+    norm_eps: float
+
+    def __init__(
+        self,
+        dim: int,
+        hc_mult: int,
+        norm_eps: float,
+        sinkhorn_iters: int,
+        hc_eps: float,
+    ) -> None: ...
+
+class HyperHead(nn.Module):
+    dim: int
+    hc_mult: int
+
+    def __init__(
+        self,
+        dim: int,
+        hc_mult: int,
+        norm_eps: float,
+        hc_eps: float,
+    ) -> None: ...
+    def __call__(self, x: mx.array) -> mx.array: ...
+
+class Compressor(nn.Module):
+    dim: int
+    head_dim: int
+    rope_head_dim: int
+    compress_ratio: int
+    overlap: bool
+    wkv_gate: nn.Linear
+    ape: mx.array
+    norm: nn.RMSNorm
+    rope: DeepseekV4RoPE
+
+    def __init__(
+        self,
+        dim: int,
+        compress_ratio: int,
+        head_dim: int,
+        rope_head_dim: int,
+        rms_norm_eps: float,
+        rope: DeepseekV4RoPE,
+    ) -> None: ...
+    def __call__(
+        self,
+        x: mx.array,
+        cache: "DeepseekV4Cache",
+        offset: Any,
+        key: str = ...,
+    ) -> mx.array: ...
+
+class Indexer(nn.Module):
+    def __init__(
+        self,
+        args: ModelArgs,
+        compress_ratio: int,
+        rope: DeepseekV4RoPE,
+    ) -> None: ...
+
+class _CompressorBranch:
+    buffer_kv: Optional[mx.array]
+    buffer_gate: Optional[mx.array]
+    prev_kv: Optional[mx.array]
+    prev_gate: Optional[mx.array]
+    pool: Optional[mx.array]
+    buffer_lengths: Optional[List[int]]
+    pool_lengths: Optional[List[int]]
+    buffer_count: int
+    _new_pool_lengths: Optional[List[int]]
+
+    def __init__(self) -> None: ...
+
+class DeepseekV4Cache:
+    local: RotatingKVCache
+    offset: int
+    keys: Optional[mx.array]
+    values: Optional[mx.array]
+    state: Any
+    meta_state: Any
+    nbytes: int
+    _branches: Dict[str, _CompressorBranch]
+    _pending_lengths: Optional[List[int]]
+
+    def __init__(self, sliding_window: int) -> None: ...
+    def update_and_fetch(
+        self, keys: mx.array, values: mx.array
+    ) -> tuple[mx.array, mx.array]: ...
+    def is_trimmable(self) -> bool: ...
+    def trim(self, n: int) -> int: ...
+    def empty(self) -> bool: ...
+    def size(self) -> int: ...
+    def prepare(
+        self,
+        *,
+        left_padding: Optional[List[int]] = None,
+        lengths: Optional[List[int]] = None,
+        right_padding: Optional[List[int]] = None,
+    ) -> None: ...
+    def finalize(self) -> None: ...
+    def filter(self, batch_indices: mx.array) -> None: ...
+    def extend(self, other: "DeepseekV4Cache") -> None: ...
+    def extract(self, idx: int) -> "DeepseekV4Cache": ...
+    @classmethod
+    def merge(cls, caches: List["DeepseekV4Cache"]) -> "DeepseekV4Cache": ...
+
+class V4Attention(nn.Module):
+    args: ModelArgs
+    layer_id: int
+    dim: int
+    n_heads: int
+    head_dim: int
+    rope_head_dim: int
+    nope_head_dim: int
+    n_groups: int
+    q_lora_rank: int
+    o_lora_rank: int
+    window: int
+    eps: float
+    scale: float
+    compress_ratio: int
+    wqkv_a: nn.Linear
+    q_norm: nn.RMSNorm
+    wq_b: nn.Linear
+    kv_norm: nn.RMSNorm
+    attn_sink: mx.array
+    wo_a: nn.Linear
+    wo_b: nn.Linear
+    rope: DeepseekV4RoPE
+    compressor: Compressor
+    indexer: Indexer
+
+    def __init__(self, args: ModelArgs, layer_id: int) -> None: ...
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array: ...
+
+class DeepseekV4MLP(nn.Module):
+    gate_proj: nn.Linear
+    up_proj: nn.Linear
+    down_proj: nn.Linear
+
+    def __init__(
+        self,
+        hidden_size: int,
+        intermediate_size: int,
+        swiglu_limit: float = 0.0,
+    ) -> None: ...
+    def __call__(self, x: mx.array) -> mx.array: ...
+
+class MoEGate(nn.Module):
+    weight: mx.array
+
+    def __init__(self, args: ModelArgs, layer_id: int) -> None: ...
+    def __call__(
+        self, x: mx.array, input_ids: mx.array
+    ) -> tuple[mx.array, mx.array]: ...
+
+class DeepseekV4MoE(nn.Module):
+    num_experts_per_tok: int
+    switch_mlp: SwitchGLU
+    gate: MoEGate
+    shared_experts: DeepseekV4MLP
+
+    def __init__(self, args: ModelArgs, layer_id: int) -> None: ...
+    def __call__(self, x: mx.array, input_ids: mx.array) -> mx.array: ...
+
+class DeepseekV4Block(nn.Module):
+    attn_norm: nn.RMSNorm
+    attn: V4Attention
+    hc_attn: HyperConnection
+    ffn_norm: nn.RMSNorm
+    ffn: DeepseekV4MoE
+    hc_ffn: HyperConnection
+
+    def __init__(self, args: ModelArgs, layer_id: int) -> None: ...
+    def __call__(
+        self,
+        h: mx.array,
+        cache: Optional[Any],
+        input_ids: mx.array,
+    ) -> mx.array: ...
+
+class DeepseekV4Model(nn.Module):
+    args: ModelArgs
+    vocab_size: int
+    embed_tokens: nn.Embedding
+    layers: list[DeepseekV4Block]
+    norm: nn.RMSNorm
+    hc_head: HyperHead
+
+    def __init__(self, args: ModelArgs) -> None: ...
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[List[Any]] = None,
+    ) -> mx.array: ...
+
+class Model(nn.Module):
+    args: ModelArgs
+    model_type: str
+    model: DeepseekV4Model
+    lm_head: nn.Linear
+
+    def __init__(self, args: ModelArgs) -> None: ...
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[List[Any]] = None,
+    ) -> mx.array: ...
+    def sanitize(self, weights: dict[str, Any]) -> dict[str, Any]: ...
+    def make_cache(self) -> list[RotatingKVCache | DeepseekV4Cache]: ...
+    @property
+    def layers(self) -> list[DeepseekV4Block]: ...
@@ -0,0 +1,31 @@
+from dataclasses import dataclass
+from typing import Any, Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from . import gemma4_text
+from .base import BaseModelArgs
+from .cache import KVCache, RotatingKVCache
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    text_config: Optional[dict[str, Any]]
+    vocab_size: int
+
+    def __post_init__(self) -> None: ...
+
+class Model(nn.Module):
+    args: ModelArgs
+    model_type: str
+    language_model: gemma4_text.Model
+
+    def __init__(self, args: ModelArgs) -> None: ...
+    def __call__(self, *args: Any, **kwargs: Any) -> mx.array: ...
+    def sanitize(self, weights: dict[str, Any]) -> dict[str, Any]: ...
+    @property
+    def layers(self) -> list[gemma4_text.DecoderLayer]: ...
+    @property
+    def quant_predicate(self) -> Any: ...
+    def make_cache(self) -> list[KVCache | RotatingKVCache]: ...
@@ -0,0 +1,179 @@
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .base import BaseModelArgs
+from .cache import KVCache, RotatingKVCache
+from .switch_layers import SwitchGLU
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    hidden_size: int
+    num_hidden_layers: int
+    intermediate_size: int
+    num_attention_heads: int
+    head_dim: int
+    global_head_dim: int
+    global_partial_rotary_factor: float
+    rms_norm_eps: float
+    vocab_size: int
+    vocab_size_per_layer_input: int
+    num_key_value_heads: int
+    num_global_key_value_heads: Optional[int]
+    num_kv_shared_layers: int
+    pad_token_id: int
+    hidden_size_per_layer_input: int
+    rope_traditional: bool
+    partial_rotary_factor: float
+    rope_parameters: Optional[Dict[str, Any]]
+    sliding_window: int
+    sliding_window_pattern: int
+    max_position_embeddings: int
+    attention_k_eq_v: bool
+    final_logit_softcapping: float
+    use_double_wide_mlp: bool
+    enable_moe_block: bool
+    num_experts: Optional[int]
+    top_k_experts: Optional[int]
+    moe_intermediate_size: Optional[int]
+    layer_types: Optional[List[str]]
+    tie_word_embeddings: bool
+
+    def __post_init__(self) -> None: ...
+
+class MLP(nn.Module):
+    gate_proj: nn.Linear
+    down_proj: nn.Linear
+    up_proj: nn.Linear
+
+    def __init__(self, config: ModelArgs, layer_idx: int = 0) -> None: ...
+    def __call__(self, x: mx.array) -> mx.array: ...
+
+class Router(nn.Module):
+    proj: nn.Linear
+    scale: mx.array
+    per_expert_scale: mx.array
+
+    def __init__(self, config: ModelArgs) -> None: ...
+    def __call__(self, x: mx.array) -> tuple[mx.array, mx.array]: ...
+
+class Experts(nn.Module):
+    switch_glu: SwitchGLU
+
+    def __init__(self, config: ModelArgs) -> None: ...
+    def __call__(
+        self, x: mx.array, top_k_indices: mx.array, top_k_weights: mx.array
+    ) -> mx.array: ...
+
+class Attention(nn.Module):
+    layer_idx: int
+    layer_type: str
+    is_sliding: bool
+    head_dim: int
+    n_heads: int
+    n_kv_heads: int
+    use_k_eq_v: bool
+    scale: float
+    q_proj: nn.Linear
+    k_proj: nn.Linear
+    v_proj: nn.Linear
+    o_proj: nn.Linear
+    q_norm: nn.Module
+    k_norm: nn.Module
+    v_norm: nn.Module
+    rope: nn.Module
+
+    def __init__(self, config: ModelArgs, layer_idx: int) -> None: ...
+    def __call__(self, *args: Any, **kwargs: Any) -> Any: ...
+
+class DecoderLayer(nn.Module):
+    layer_idx: int
+    layer_type: str
+    self_attn: Attention
+    mlp: MLP
+    enable_moe: bool
+    router: Router
+    experts: Experts
+    input_layernorm: nn.Module
+    post_attention_layernorm: nn.Module
+    pre_feedforward_layernorm: nn.Module
+    post_feedforward_layernorm: nn.Module
+    post_feedforward_layernorm_1: nn.Module
+    post_feedforward_layernorm_2: nn.Module
+    pre_feedforward_layernorm_2: nn.Module
+    hidden_size_per_layer_input: int
+    per_layer_input_gate: Optional[nn.Linear]
+    per_layer_projection: Optional[nn.Linear]
+    post_per_layer_input_norm: Optional[nn.Module]
+    layer_scalar: mx.array
+
+    def __init__(self, config: ModelArgs, layer_idx: int) -> None: ...
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = ...,
+        cache: Optional[Any] = ...,
+        per_layer_input: Optional[mx.array] = ...,
+        shared_kv: Optional[tuple[mx.array, mx.array]] = ...,
+        offset: Optional[mx.array] = ...,
+    ) -> tuple[mx.array, tuple[mx.array, mx.array], mx.array]: ...
+
+class Gemma4TextModel(nn.Module):
+    config: ModelArgs
+    vocab_size: int
+    window_size: int
+    sliding_window_pattern: int
+    num_hidden_layers: int
+    embed_tokens: nn.Embedding
+    embed_scale: float
+    layers: list[DecoderLayer]
+    norm: nn.Module
+    hidden_size_per_layer_input: int
+    embed_tokens_per_layer: Optional[nn.Embedding]
+    per_layer_model_projection: Optional[nn.Linear]
+    per_layer_projection_norm: Optional[nn.Module]
+    previous_kvs: list[int]
+
+    def __init__(self, config: ModelArgs) -> None: ...
+    def __call__(
+        self,
+        inputs: Optional[mx.array] = ...,
+        cache: Optional[list[Any]] = ...,
+        input_embeddings: Optional[mx.array] = ...,
+        per_layer_inputs: Optional[mx.array] = ...,
+    ) -> mx.array: ...
+    def _get_per_layer_inputs(
+        self,
+        input_ids: Optional[mx.array],
+        input_embeddings: Optional[mx.array] = ...,
+    ) -> mx.array: ...
+    def _project_per_layer_inputs(
+        self,
+        input_embeddings: mx.array,
+        per_layer_inputs: Optional[mx.array] = ...,
+    ) -> mx.array: ...
+    def _make_masks(self, h: mx.array, cache: list[Any]) -> list[Any]: ...
+
+class Model(nn.Module):
+    args: ModelArgs
+    model_type: str
+    model: Gemma4TextModel
+    final_logit_softcapping: float
+    tie_word_embeddings: bool
+    lm_head: nn.Linear
+
+    def __init__(self, args: ModelArgs) -> None: ...
+    def __call__(self, *args: Any, **kwargs: Any) -> mx.array: ...
+    def sanitize(self, weights: dict[str, Any]) -> dict[str, Any]: ...
+    @property
+    def layers(self) -> list[DecoderLayer]: ...
+    @property
+    def head_dim(self) -> int: ...
+    @property
+    def n_kv_heads(self) -> int: ...
+    @property
+    def quant_predicate(self) -> Any: ...
+    def make_cache(self) -> list[KVCache | RotatingKVCache]: ...
@@ -0,0 +1,103 @@
+"""Type stubs for mlx_lm.models.gpt_oss"""
+
+from dataclasses import dataclass
+from typing import Any, List, Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .base import BaseModelArgs
+from .cache import KVCache
+from .switch_layers import SwitchGLU
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    hidden_size: int
+    intermediate_size: int
+    num_hidden_layers: int
+    num_attention_heads: int
+    num_key_value_heads: int
+    num_local_experts: int
+    num_experts_per_tok: int
+    vocab_size: int
+    rms_norm_eps: float
+    sliding_window: int
+    layer_types: Optional[List[str]]
+
+def mlx_topk(a: mx.array, k: int, axis: int = -1) -> tuple[mx.array, mx.array]: ...
+
+class AttentionBlock(nn.Module):
+    head_dim: int
+    num_attention_heads: int
+    num_key_value_heads: int
+    num_key_value_groups: int
+    sinks: mx.array
+    q_proj: nn.Linear
+    k_proj: nn.Linear
+    v_proj: nn.Linear
+    o_proj: nn.Linear
+    sm_scale: float
+    rope: nn.Module
+
+    def __init__(self, config: ModelArgs) -> None: ...
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array: ...
+
+class TransformerBlock(nn.Module):
+    self_attn: AttentionBlock
+    mlp: MLPBlock
+
+    def __init__(self, config: ModelArgs) -> None: ...
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array: ...
+
+class MLPBlock(nn.Module):
+    hidden_size: int
+    num_local_experts: int
+    num_experts_per_tok: int
+    experts: SwitchGLU
+    router: nn.Linear
+    sharding_group: Optional[mx.distributed.Group]
+
+    def __init__(self, config: ModelArgs) -> None: ...
+    def __call__(self, x: mx.array) -> mx.array: ...
+
+class GptOssMoeModel(nn.Module):
+    embed_tokens: nn.Embedding
+    norm: nn.RMSNorm
+    layer_types: List[str]
+    layers: list[TransformerBlock]
+    window_size: int
+    swa_idx: int
+    ga_idx: int
+
+    def __init__(self, args: ModelArgs) -> None: ...
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[Any] = None,
+    ) -> mx.array: ...
+
+class Model(nn.Module):
+    model_type: str
+    model: GptOssMoeModel
+    lm_head: nn.Linear
+
+    def __init__(self, args: ModelArgs) -> None: ...
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[Any] = None,
+    ) -> mx.array: ...
+    @property
+    def layers(self) -> list[nn.Module]: ...
+    def make_cache(self) -> list[KVCache]: ...
@@ -0,0 +1,94 @@
+"""Type stubs for mlx_lm.models.minimax"""
+
+from dataclasses import dataclass
+from typing import Any, Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .base import BaseModelArgs
+from .switch_layers import SwitchGLU
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    hidden_size: int
+    intermediate_size: int
+    num_hidden_layers: int
+    num_attention_heads: int
+    num_key_value_heads: int
+    num_local_experts: int
+    num_experts_per_tok: int
+    max_position_embeddings: int
+
+class MiniMaxAttention(nn.Module):
+    num_heads: int
+    num_attention_heads: int
+    num_key_value_heads: int
+    head_dim: int
+    scale: float
+    q_proj: nn.Linear
+    k_proj: nn.Linear
+    v_proj: nn.Linear
+    o_proj: nn.Linear
+    q_norm: nn.Module
+    k_norm: nn.Module
+    rope: nn.Module
+
+    def __init__(self, args: ModelArgs) -> None: ...
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array: ...
+
+class MiniMaxSparseMoeBlock(nn.Module):
+    num_experts_per_tok: int
+    gate: nn.Linear
+    switch_mlp: SwitchGLU
+    e_score_correction_bias: mx.array
+    sharding_group: Optional[mx.distributed.Group]
+
+    def __init__(self, args: ModelArgs) -> None: ...
+    def __call__(self, x: mx.array) -> mx.array: ...
+
+class MiniMaxDecoderLayer(nn.Module):
+    self_attn: MiniMaxAttention
+    block_sparse_moe: MiniMaxSparseMoeBlock
+    input_layernorm: nn.RMSNorm
+    post_attention_layernorm: nn.RMSNorm
+
+    def __init__(self, args: ModelArgs) -> None: ...
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array: ...
+
+class MiniMaxModel(nn.Module):
+    embed_tokens: nn.Embedding
+    layers: list[MiniMaxDecoderLayer]
+    norm: nn.RMSNorm
+
+    def __init__(self, args: ModelArgs) -> None: ...
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[Any] = None,
+    ) -> mx.array: ...
+
+class Model(nn.Module):
+    model_type: str
+    model: MiniMaxModel
+    lm_head: nn.Linear
+
+    def __init__(self, args: ModelArgs) -> None: ...
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[Any] = None,
+    ) -> mx.array: ...
+    @property
+    def layers(self) -> list[MiniMaxDecoderLayer]: ...
@@ -92,6 +92,15 @@ class NemotronHAttention(nn.Module):
        cache: Optional[KVCache] = None,
    ) -> mx.array: ...

+class MoEGate(nn.Module):
+    config: ModelArgs
+    top_k: int
+    norm_topk_prob: bool
+    weight: mx.array
+
+    def __init__(self, config: ModelArgs) -> None: ...
+    def __call__(self, x: mx.array) -> tuple[mx.array, mx.array]: ...
+
 class NemotronHMLP(nn.Module):
    up_proj: nn.Linear
    down_proj: nn.Linear
@@ -102,9 +111,14 @@ class NemotronHMLP(nn.Module):
    def __call__(self, x: mx.array) -> mx.array: ...

 class NemotronHMoE(nn.Module):
+    config: ModelArgs
    num_experts_per_tok: int
+    moe_latent_size: Optional[int]
    switch_mlp: SwitchMLP
+    gate: MoEGate
    shared_experts: NemotronHMLP
+    fc1_latent_proj: nn.Linear
+    fc2_latent_proj: nn.Linear

    def __init__(self, config: ModelArgs) -> None: ...
    def __call__(self, x: mx.array) -> mx.array: ...
@@ -71,6 +71,7 @@ class Qwen3NextAttention(nn.Module):
 class Qwen3NextSparseMoeBlock(nn.Module):
    norm_topk_prob: bool
    num_experts: int
+    num_experts_per_tok: int
    top_k: int
    gate: nn.Linear
    switch_mlp: SwitchGLU
@@ -48,6 +48,10 @@ def make_logits_processors(
    logit_bias: Optional[Dict[int, float]] = ...,
    repetition_penalty: Optional[float] = ...,
    repetition_context_size: Optional[int] = ...,
+    presence_penalty: Optional[float] = ...,
+    presence_context_size: Optional[int] = ...,
+    frequency_penalty: Optional[float] = ...,
+    frequency_context_size: Optional[int] = ...,
 ) -> list[Callable[[mx.array, mx.array], mx.array]]:
    """
    Make logits processors for use with ``generate_step``.
@@ -117,6 +117,8 @@ class TokenizerWrapper:
    think_end: str | None
    think_start_id: int | None
    think_end_id: int | None
+    think_start_tokens: list[int] | None
+    think_end_tokens: list[int] | None

    def __init__(
        self,
@@ -0,0 +1,50 @@
+"""
+This type stub file was generated by pyright.
+"""
+
+import mlx.nn as nn
+
+class DoRALinear(nn.Module):
+    @staticmethod
+    def from_base(
+        linear: nn.Linear, r: int = ..., dropout: float = ..., scale: float = ...
+    ):  # -> DoRALinear:
+        ...
+    def fuse(self, dequantize: bool = ...):  # -> QuantizedLinear | Linear:
+        ...
+    def __init__(
+        self,
+        input_dims: int,
+        output_dims: int,
+        r: int = ...,
+        dropout: float = ...,
+        scale: float = ...,
+        bias: bool = ...,
+    ) -> None: ...
+    def set_linear(self, linear):  # -> None:
+        """
+        Set the self.linear layer and recompute self.m.
+        """
+        ...
+
+    def __call__(self, x): ...
+
+class DoRAEmbedding(nn.Module):
+    def from_base(
+        embedding: nn.Embedding, r: int = ..., dropout: float = ..., scale: float = ...
+    ):  # -> DoRAEmbedding:
+        ...
+    def fuse(self, dequantize: bool = ...):  # -> Embedding:
+        ...
+    def __init__(
+        self,
+        num_embeddings: int,
+        dims: int,
+        r: int = ...,
+        dropout: float = ...,
+        scale: float = ...,
+    ) -> None: ...
+    def set_embedding(self, embedding: nn.Module):  # -> None:
+        ...
+    def __call__(self, x): ...
+    def as_linear(self, x): ...
@@ -0,0 +1,66 @@
+"""
+This type stub file was generated by pyright.
+"""
+
+import mlx.nn as nn
+
+class LoRALinear(nn.Module):
+    @staticmethod
+    def from_base(
+        linear: nn.Linear, r: int = ..., dropout: float = ..., scale: float = ...
+    ):  # -> LoRALinear:
+        ...
+    def fuse(self, dequantize: bool = ...):  # -> QuantizedLinear | Linear:
+        ...
+    def __init__(
+        self,
+        input_dims: int,
+        output_dims: int,
+        r: int = ...,
+        dropout: float = ...,
+        scale: float = ...,
+        bias: bool = ...,
+    ) -> None: ...
+    def __call__(self, x):  # -> array:
+        ...
+
+class LoRASwitchLinear(nn.Module):
+    @staticmethod
+    def from_base(
+        linear: nn.Module, r: int = ..., dropout: float = ..., scale: float = ...
+    ):  # -> LoRASwitchLinear:
+        ...
+    def fuse(self, dequantize: bool = ...):  # -> QuantizedSwitchLinear | SwitchLinear:
+        ...
+    def __init__(
+        self,
+        input_dims: int,
+        output_dims: int,
+        num_experts: int,
+        r: int = ...,
+        dropout: float = ...,
+        scale: float = ...,
+        bias: bool = ...,
+    ) -> None: ...
+    def __call__(self, x, indices, sorted_indices=...): ...
+
+class LoRAEmbedding(nn.Module):
+    @staticmethod
+    def from_base(
+        embedding: nn.Embedding, r: int = ..., dropout: float = ..., scale: float = ...
+    ):  # -> LoRAEmbedding:
+        ...
+    def fuse(self, dequantize: bool = ...):  # -> QuantizedEmbedding | Embedding:
+        ...
+    def __init__(
+        self,
+        num_embeddings: int,
+        dims: int,
+        r: int = ...,
+        dropout: float = ...,
+        scale: float = ...,
+    ) -> None: ...
+    def __call__(self, x):  # -> array:
+        ...
+    def as_linear(self, x):  # -> array:
+        ...
@@ -0,0 +1,57 @@
+"""
+This type stub file was generated by pyright.
+"""
+
+import mlx.nn as nn
+from typing import Dict
+
+def build_schedule(schedule_config: Dict):  # -> Any:
+    """
+    Build a learning rate schedule from the given config.
+    """
+    ...
+
+def linear_to_lora_layers(
+    model: nn.Module, num_layers: int, config: Dict, use_dora: bool = ...
+):  # -> None:
+    """
+    Convert some of the models linear layers to lora layers.
+
+    Args:
+        model (nn.Module): The neural network model.
+        num_layers (int): The number of blocks to convert to lora layers
+        starting from the last layer.
+        config (dict): More configuration parameters for LoRA, including the
+          rank, scale, and optional layer keys.
+        use_dora (bool): If True, uses DoRA instead of LoRA.
+          Default: ``False``
+    """
+    ...
+
+def load_adapters(model: nn.Module, adapter_path: str) -> nn.Module:
+    """
+    Load any fine-tuned adapters / layers.
+
+    Args:
+        model (nn.Module): The neural network model.
+        adapter_path (str): Path to the adapter configuration file.
+
+    Returns:
+        nn.Module: The updated model with LoRA layers applied.
+    """
+    ...
+
+def remove_lora_layers(model: nn.Module) -> nn.Module:
+    """
+    Remove the LoRA layers from the model.
+
+    Args:
+        model (nn.Module): The model with LoRA layers.
+
+    Returns:
+        nn.Module: The model without LoRA layers.
+    """
+    ...
+
+def print_trainable_parameters(model):  # -> None:
+    ...
@@ -0,0 +1,12 @@
+from typing import Any
+
+def get_message_json(
+    model_name: str,
+    prompt: str,
+    role: str = "user",
+    skip_image_token: bool = False,
+    skip_audio_token: bool = False,
+    num_images: int = 0,
+    num_audios: int = 0,
+    **kwargs: Any,
+) -> dict[str, Any]: ...
@@ -0,0 +1,15 @@
+from pathlib import Path
+from typing import Any
+
+class ImageProcessor:
+    def preprocess(
+        self, images: list[dict[str, Any]], **kwargs: Any
+    ) -> dict[str, Any]: ...
+    def __call__(self, **kwargs: Any) -> dict[str, Any]: ...
+
+def load_image_processor(
+    model_path: str | Path, **kwargs: Any
+) -> ImageProcessor | None: ...
+def load_processor(
+    model_path: str | Path, add_detokenizer: bool = ..., **kwargs: Any
+) -> ImageProcessor: ...
@@ -0,0 +1,8 @@
+from typing import Any, Self
+
+class safe_open:
+    def __init__(self, filename: str, framework: str = "pt") -> None: ...
+    def __enter__(self) -> Self: ...
+    def __exit__(self, *args: Any) -> None: ...
+    def keys(self) -> list[str]: ...
+    def get_tensor(self, name: str) -> Any: ...
@@ -1,7 +1,8 @@
 {
    "recommendations": [
        "detachhead.basedpyright",
-        "ms-python.python"
+        "ms-python.python",
+        "jnoortheen.nix-ide"
    ],
    "unwantedRecommendations": [
        "ms-python.vscode-pylance",
@@ -1,3 +1,32 @@
 {
-    "basedpyright.importStrategy": "fromEnvironment"
+    "files.associations": {
+        "*.nix": "nix",
+    },
+    "nix.enableLanguageServer": true,
+    "nix.serverPath": "nixd",
+    "nix.serverSettings": {
+        "nixd": {
+        "formatting": {
+            "command": ["nixpkgs-fmt"]
+        },
+        "nixpkgs": {
+            "expr": "(builtins.getFlake \"path:${workspaceFolder}\").currentSystem.config._module.args.pkgs"
+        },
+        "options": {
+            "flake-parts": {
+                "expr": "(builtins.getFlake \"path:${workspaceFolder}\").debug.options"
+            },
+            "flake-parts-perSystem": {
+                "expr": "(builtins.getFlake \"path:${workspaceFolder}\").currentSystem.options"
+            }
+        }
+      }
+    },
+    "[nix]": {
+        "editor.defaultFormatter": "jnoortheen.nix-ide"
+    },
+    
+    "python.defaultInterpreterPath": "${workspaceFolder}/.venv/bin/python",
+    "basedpyright.analysis.configFilePath": "${workspaceFolder}/pyproject.toml",
+    "basedpyright.importStrategy": "fromEnvironment",
 }
@@ -19,8 +19,8 @@ To run EXO from source:
 - [macmon](https://github.com/vladkens/macmon) (for hardware monitoring on Apple Silicon)
  Use the pinned fork revision used by this repo instead of Homebrew `macmon`.
  ```bash
-  cargo install --git https://github.com/swiftraccoon/macmon \
-    --rev 9154d234f763fbeffdcb4135d0bbbaf80609699b \
+  cargo install --git https://github.com/vladkens/macmon \
+    --rev a1cd06b6cc0d5e61db24fd8832e74cd992097a7d \
    macmon \
    --force
  ```
@@ -112,8 +112,8 @@ Then restart the Nix daemon: `sudo launchctl kickstart -k system/org.nixos.nix-d
  Homebrew `macmon 0.6.1` still crashes on Apple M5.

  ```bash
-  cargo install --git https://github.com/swiftraccoon/macmon \
-    --rev 9154d234f763fbeffdcb4135d0bbbaf80609699b \
+  cargo install --git https://github.com/vladkens/macmon \
+    --rev a1cd06b6cc0d5e61db24fd8832e74cd992097a7d \
    macmon \
    --force
  ```
@@ -295,8 +295,9 @@ exo supports several environment variables for configuration:

 | Variable | Description | Default |
 |----------|-------------|---------|
-| `EXO_MODELS_PATH` | Colon-separated paths to search for pre-downloaded models (e.g., on NFS mounts or shared storage) | None |
-| `EXO_MODELS_DIR` | Directory where exo downloads and stores models | `~/.local/share/exo/models` (Linux) or `~/.exo/models` (macOS) |
+| `EXO_DEFAULT_MODELS_DIR` | Default directory for model downloads and caches. Always first in the writable dirs list. | `~/.local/share/exo/models` (Linux) or `~/.exo/models` (macOS) |
+| `EXO_MODELS_DIRS` | Colon-separated additional writable directories for model downloads. Checked in order after the default; first with enough free space is used. | None |
+| `EXO_MODELS_READ_ONLY_DIRS` | Colon-separated read-only directories to search for pre-downloaded models (e.g., NFS mounts, shared storage). Models here cannot be deleted. | None |
 | `EXO_OFFLINE` | Run without internet connection (uses only local models) | `false` |
 | `EXO_ENABLE_IMAGE_MODELS` | Enable image model support | `false` |
 | `EXO_LIBP2P_NAMESPACE` | Custom namespace for cluster isolation | None |
@@ -306,8 +307,11 @@ exo supports several environment variables for configuration:
 **Example usage:**

 ```bash
-# Use pre-downloaded models from NFS mount
-EXO_MODELS_PATH=/mnt/nfs/models:/opt/ai-models uv run exo
+# Use pre-downloaded models from NFS mount (read-only)
+EXO_MODELS_READ_ONLY_DIRS=/mnt/nfs/models:/opt/ai-models uv run exo
+
+# Download models to an external SSD (falls back to default dir if full)
+EXO_MODELS_DIRS=/Volumes/ExternalSSD/exo-models uv run exo

 # Run in offline mode
 EXO_OFFLINE=true uv run exo
@@ -16,22 +16,13 @@ struct ContentView: View {
    @EnvironmentObject private var updater: SparkleUpdater
    @EnvironmentObject private var thunderboltBridgeService: ThunderboltBridgeService
    @EnvironmentObject private var settingsWindowController: SettingsWindowController
+    @EnvironmentObject private var bugReportWindowController: BugReportWindowController
    @State private var focusedNode: NodeViewModel?
    @State private var deletingInstanceIDs: Set<String> = []
    @State private var showAllNodes = false
    @State private var showAllInstances = false
    @State private var baseURLCopied = false
    @State private var showAdvanced = false
-    @State private var showDebugInfo = false
-    private enum BugReportPhase: Equatable {
-        case idle
-        case prompting
-        case sending(String)
-        case success(String)
-        case failure(String)
-    }
-    @State private var bugReportPhase: BugReportPhase = .idle
-    @State private var bugReportUserDescription: String = ""
    @State private var uninstallInProgress = false
    @State private var pendingNamespace: String = ""
    @State private var pendingHFToken: String = ""
@@ -294,6 +285,13 @@ struct ContentView: View {
            ) {
                updater.checkForUpdates()
            }
+            HoverButton(
+                title: "Share Bug Report…",
+                tint: .primary,
+                trailingSystemImage: "ladybug"
+            ) {
+                bugReportWindowController.open()
+            }
            .padding(.bottom, 8)
            HoverButton(title: "Quit", tint: .secondary) {
                controller.stop()
@@ -477,40 +475,6 @@ struct ContentView: View {
        }
    }

-    private var debugSection: some View {
-        VStack(alignment: .leading, spacing: 4) {
-            HoverButton(
-                title: "Debug Info",
-                tint: .primary,
-                trailingSystemImage: showDebugInfo ? "chevron.up" : "chevron.down",
-                small: true
-            ) {
-                showDebugInfo.toggle()
-            }
-            if showDebugInfo {
-                VStack(alignment: .leading, spacing: 4) {
-                    Text("Version: \(buildTag)")
-                        .font(.caption2)
-                        .foregroundColor(.secondary)
-                    Text("Commit: \(buildCommit)")
-                        .font(.caption2)
-                        .foregroundColor(.secondary)
-                    Text(thunderboltStatusText)
-                        .font(.caption2)
-                        .foregroundColor(thunderboltStatusColor)
-                    clusterThunderboltBridgeView
-                    interfaceIpList
-                    rdmaStatusView
-                    sendBugReportButton
-                        .padding(.top, 6)
-                }
-                .padding(.leading, 8)
-                .transition(.opacity)
-            }
-        }
-        .animation(.easeInOut(duration: 0.25), value: showDebugInfo)
-    }
-
    private var rdmaStatusView: some View {
        let rdmaStatuses = stateService.latestSnapshot?.nodeRdmaCtl ?? [:]
        let localNodeId = stateService.localNodeId
@@ -559,118 +523,6 @@ struct ContentView: View {
        }
    }

-    private var sendBugReportButton: some View {
-        VStack(alignment: .leading, spacing: 6) {
-            switch bugReportPhase {
-            case .idle:
-                Button {
-                    bugReportPhase = .prompting
-                    bugReportUserDescription = ""
-                } label: {
-                    HStack {
-                        Text("Send Bug Report")
-                            .font(.caption)
-                            .fontWeight(.semibold)
-                        Spacer()
-                    }
-                    .padding(.vertical, 6)
-                    .padding(.horizontal, 8)
-                    .background(
-                        RoundedRectangle(cornerRadius: 6)
-                            .fill(Color.accentColor.opacity(0.12))
-                    )
-                }
-                .buttonStyle(.plain)
-
-            case .prompting:
-                VStack(alignment: .leading, spacing: 6) {
-                    Text("What's the issue? (optional)")
-                        .font(.caption2)
-                        .foregroundColor(.secondary)
-                    TextEditor(text: $bugReportUserDescription)
-                        .font(.caption2)
-                        .frame(height: 60)
-                        .overlay(
-                            RoundedRectangle(cornerRadius: 4)
-                                .stroke(Color.secondary.opacity(0.3), lineWidth: 1)
-                        )
-                    HStack(spacing: 8) {
-                        Button("Send") {
-                            Task {
-                                await sendBugReport()
-                            }
-                        }
-                        .font(.caption2)
-                        .buttonStyle(.borderedProminent)
-                        .controlSize(.small)
-                        Button("Cancel") {
-                            bugReportPhase = .idle
-                        }
-                        .font(.caption2)
-                        .buttonStyle(.bordered)
-                        .controlSize(.small)
-                    }
-                }
-                .padding(8)
-                .background(
-                    RoundedRectangle(cornerRadius: 6)
-                        .fill(Color.accentColor.opacity(0.06))
-                )
-
-            case .sending(let message):
-                HStack(spacing: 6) {
-                    ProgressView()
-                        .scaleEffect(0.6)
-                    Text(message)
-                        .font(.caption2)
-                        .foregroundColor(.secondary)
-                }
-
-            case .success(let message):
-                VStack(alignment: .leading, spacing: 6) {
-                    Text(message)
-                        .font(.caption2)
-                        .foregroundColor(.secondary)
-                        .fixedSize(horizontal: false, vertical: true)
-                    Button {
-                        openGitHubIssue()
-                    } label: {
-                        HStack(spacing: 4) {
-                            Image(systemName: "arrow.up.right.square")
-                                .imageScale(.small)
-                            Text("Create GitHub Issue")
-                                .font(.caption2)
-                        }
-                    }
-                    .buttonStyle(.bordered)
-                    .controlSize(.small)
-                    Button("Done") {
-                        bugReportPhase = .idle
-                        bugReportUserDescription = ""
-                    }
-                    .font(.caption2)
-                    .buttonStyle(.plain)
-                    .foregroundColor(.secondary)
-                }
-
-            case .failure(let message):
-                VStack(alignment: .leading, spacing: 4) {
-                    Text(message)
-                        .font(.caption2)
-                        .foregroundColor(.red)
-                        .fixedSize(horizontal: false, vertical: true)
-                    Button("Dismiss") {
-                        bugReportPhase = .idle
-                    }
-                    .font(.caption2)
-                    .buttonStyle(.plain)
-                    .foregroundColor(.secondary)
-                }
-            }
-        }
-        .animation(.easeInOut(duration: 0.2), value: bugReportPhase)
-    }
-
    private var processToggleBinding: Binding<Bool> {
        Binding(
            get: {
@@ -711,61 +563,6 @@ struct ContentView: View {
        )
    }

-    private func sendBugReport() async {
-        bugReportPhase = .sending("Collecting logs...")
-        let service = BugReportService()
-        let description = bugReportUserDescription.trimmingCharacters(in: .whitespacesAndNewlines)
-        do {
-            let outcome = try await service.sendReport(
-                isManual: true,
-                userDescription: description.isEmpty ? nil : description
-            )
-            if outcome.success {
-                bugReportPhase = .success(outcome.message)
-            } else {
-                bugReportPhase = .failure(outcome.message)
-            }
-        } catch {
-            bugReportPhase = .failure(error.localizedDescription)
-        }
-    }
-
-    private func openGitHubIssue() {
-        let description = bugReportUserDescription.trimmingCharacters(in: .whitespacesAndNewlines)
-
-        var bodyParts: [String] = []
-        bodyParts.append("## Describe the bug")
-        bodyParts.append("")
-        if !description.isEmpty {
-            bodyParts.append(description)
-        } else {
-            bodyParts.append("A clear and concise description of what the bug is.")
-        }
-        bodyParts.append("")
-        bodyParts.append("## Environment")
-        bodyParts.append("")
-        bodyParts.append("- macOS Version: \(ProcessInfo.processInfo.operatingSystemVersionString)")
-        bodyParts.append("- EXO Version: \(buildTag) (\(buildCommit))")
-        bodyParts.append("")
-        bodyParts.append("## Additional context")
-        bodyParts.append("")
-        bodyParts.append("A bug report with diagnostic logs was submitted via the app.")
-
-        let body = bodyParts.joined(separator: "\n")
-
-        var components = URLComponents(string: "https://github.com/exo-explore/exo/issues/new")!
-        components.queryItems = [
-            URLQueryItem(name: "template", value: "bug_report.md"),
-            URLQueryItem(name: "title", value: "[BUG] "),
-            URLQueryItem(name: "body", value: body),
-            URLQueryItem(name: "labels", value: "bug"),
-        ]
-
-        if let url = components.url {
-            NSWorkspace.shared.open(url)
-        }
-    }
-
    private func showUninstallConfirmationAlert() {
        let alert = NSAlert()
        alert.messageText = "Uninstall EXO"
@@ -848,13 +645,6 @@ struct ContentView: View {
        }
    }

-    private var buildTag: String {
-        Bundle.main.infoDictionary?["EXOBuildTag"] as? String ?? "unknown"
-    }
-
-    private var buildCommit: String {
-        Bundle.main.infoDictionary?["EXOBuildCommit"] as? String ?? "unknown"
-    }
 }

 private struct HoverButton: View {
@@ -22,6 +22,7 @@ struct EXOApp: App {
    @StateObject private var updater: SparkleUpdater
    @StateObject private var thunderboltBridgeService: ThunderboltBridgeService
    @StateObject private var settingsWindowController: SettingsWindowController
+    @StateObject private var bugReportWindowController: BugReportWindowController
    private let terminationObserver: TerminationObserver
    private let firstLaunchPopout = FirstLaunchPopout()
    private let ciContext = CIContext(options: nil)
@@ -46,6 +47,7 @@ struct EXOApp: App {
        let thunderboltBridge = ThunderboltBridgeService(clusterStateService: service)
        _thunderboltBridgeService = StateObject(wrappedValue: thunderboltBridge)
        _settingsWindowController = StateObject(wrappedValue: SettingsWindowController())
+        _bugReportWindowController = StateObject(wrappedValue: BugReportWindowController())
        enableLaunchAtLoginIfNeeded()
        // Install LaunchDaemon to disable Thunderbolt Bridge on startup (prevents network loops)
        NetworkSetupHelper.promptAndInstallIfNeeded()
@@ -66,6 +68,7 @@ struct EXOApp: App {
                .environmentObject(updater)
                .environmentObject(thunderboltBridgeService)
                .environmentObject(settingsWindowController)
+                .environmentObject(bugReportWindowController)
        } label: {
            menuBarIcon
                .onReceive(controller.$isFirstLaunchReady) { ready in
@@ -7,7 +7,27 @@ private let hfTokenKey = "EXOHFToken"
 private let hfEndpointKey = "EXOHFEndpoint"
 private let enableImageModelsKey = "EXOEnableImageModels"
 private let offlineModeKey = "EXOOfflineMode"
+private let fastSynchEnabledKey = "EXOFastSynchEnabled"
 private let onboardingCompletedKey = "EXOOnboardingCompleted"
+private let defaultModelsDirKey = "EXODefaultModelsDir"
+private let additionalModelsDirsKey = "EXOAdditionalModelsDirs"
+private let readOnlyModelsDirsKey = "EXOReadOnlyModelsDirs"
+private let customEnvironmentVariablesKey = "EXOCustomEnvironmentVariables"
+
+/// A user-defined environment variable that is injected into the exo child
+/// process at launch. Used as an escape hatch for env vars that don't have
+/// first-class typed UI in Settings.
+struct CustomEnvironmentVariable: Codable, Identifiable, Equatable {
+    var id: UUID
+    var key: String
+    var value: String
+
+    init(id: UUID = UUID(), key: String = "", value: String = "") {
+        self.id = id
+        self.key = key
+        self.value = value
+    }
+}

@MainActor
 final class ExoProcessController: ObservableObject {
@@ -78,6 +98,60 @@ final class ExoProcessController: ObservableObject {
            UserDefaults.standard.set(offlineMode, forKey: offlineModeKey)
        }
    }
+    @Published var fastSynchEnabled: Bool = {
+        if UserDefaults.standard.object(forKey: fastSynchEnabledKey) == nil {
+            return true
+        }
+        return UserDefaults.standard.bool(forKey: fastSynchEnabledKey)
+    }()
+    {
+        didSet {
+            UserDefaults.standard.set(fastSynchEnabled, forKey: fastSynchEnabledKey)
+        }
+    }
+    @Published var defaultModelsDir: String = {
+        return UserDefaults.standard.string(forKey: defaultModelsDirKey) ?? ""
+    }()
+    {
+        didSet {
+            UserDefaults.standard.set(defaultModelsDir, forKey: defaultModelsDirKey)
+        }
+    }
+    @Published var additionalModelsDirs: String = {
+        return UserDefaults.standard.string(forKey: additionalModelsDirsKey) ?? ""
+    }()
+    {
+        didSet {
+            UserDefaults.standard.set(additionalModelsDirs, forKey: additionalModelsDirsKey)
+        }
+    }
+    @Published var readOnlyModelsDirs: String = {
+        return UserDefaults.standard.string(forKey: readOnlyModelsDirsKey) ?? ""
+    }()
+    {
+        didSet {
+            UserDefaults.standard.set(readOnlyModelsDirs, forKey: readOnlyModelsDirsKey)
+        }
+    }
+    @Published var customEnvironmentVariables: [CustomEnvironmentVariable] = {
+        guard
+            let data = UserDefaults.standard.data(forKey: customEnvironmentVariablesKey),
+            let decoded = try? JSONDecoder().decode(
+                [CustomEnvironmentVariable].self, from: data
+            )
+        else {
+            return []
+        }
+        return decoded
+    }()
+    {
+        didSet {
+            guard let data = try? JSONEncoder().encode(customEnvironmentVariables) else {
+                return
+            }
+            UserDefaults.standard.set(data, forKey: customEnvironmentVariablesKey)
+        }
+    }

    /// Fires once when EXO transitions to `.running` for the very first time (fresh install).
    @Published private(set) var isFirstLaunchReady = false
@@ -291,6 +365,7 @@ final class ExoProcessController: ObservableObject {
        if offlineMode {
            environment["EXO_OFFLINE"] = "true"
        }
+        environment["EXO_FAST_SYNCH"] = fastSynchEnabled ? "true" : "false"

        var paths: [String] = []
        if let existing = environment["PATH"], !existing.isEmpty {
@@ -315,6 +390,29 @@ final class ExoProcessController: ObservableObject {
        }

        environment["PATH"] = paths.joined(separator: ":")
+
+        let trimmedDefaultModelsDir = defaultModelsDir.trimmingCharacters(in: .whitespaces)
+        if !trimmedDefaultModelsDir.isEmpty {
+            environment["EXO_DEFAULT_MODELS_DIR"] = trimmedDefaultModelsDir
+        }
+        let trimmedAdditionalModelsDirs = additionalModelsDirs.trimmingCharacters(in: .whitespaces)
+        if !trimmedAdditionalModelsDirs.isEmpty {
+            environment["EXO_MODELS_DIRS"] = trimmedAdditionalModelsDirs
+        }
+        let trimmedReadOnlyModelsDirs = readOnlyModelsDirs.trimmingCharacters(in: .whitespaces)
+        if !trimmedReadOnlyModelsDirs.isEmpty {
+            environment["EXO_MODELS_READ_ONLY_DIRS"] = trimmedReadOnlyModelsDirs
+        }
+
+        // Apply user-defined arbitrary environment variables last so that
+        // power users can override any of the typed fields above when
+        // necessary. Empty keys are ignored.
+        for variable in customEnvironmentVariables {
+            let trimmedKey = variable.key.trimmingCharacters(in: .whitespaces)
+            guard !trimmedKey.isEmpty else { continue }
+            environment[trimmedKey] = variable.value
+        }
+
        return environment
    }

@@ -9,7 +9,7 @@
 	<key>EXOBuildCommit</key>
 	<string>$(EXO_BUILD_COMMIT)</string>
 	<key>EXOBugReportPresignedUrlEndpoint</key>
-	<string>$(EXO_BUG_REPORT_PRESIGNED_URL_ENDPOINT)</string>
+	<string>https://reports.exolabs.net/presigned-urls</string>
 	<key>NSLocalNetworkUsageDescription</key>
 	<string>EXO needs local network access to discover and connect to other devices in your cluster for distributed AI inference.</string>
 	<key>NSBonjourServices</key>
@@ -17,7 +17,7 @@ final class ClusterStateService: ObservableObject {

    init(
        baseURL: URL = URL(string: "http://127.0.0.1:52415")!,
-        session: URLSession = .shared
+        session: URLSession = ClusterStateService.makeNonCachingSession()
    ) {
        self.baseURL = baseURL
        self.endpoint = baseURL.appendingPathComponent("state")
@@ -27,6 +27,23 @@ final class ClusterStateService: ObservableObject {
        self.decoder = decoder
    }

+    /// `URLSession.shared` carries an on-disk `URLCache` that persists every
+    /// response body under `~/Library/Caches/exolabs.EXO/`. We poll `/state`
+    /// at 2 Hz from `startPolling`, so leaving the shared cache attached
+    /// dirties ~500–620 KB/sec of file-backed memory and trips macOS's
+    /// per-process `disk writes` resource limit (microstackshot reports
+    /// observed on M3 Ultra producing GBs of cached responses per hour).
+    /// Cluster-state polling responses are time-sensitive and small; they
+    /// gain nothing from being cached on disk. Use an ephemeral session
+    /// with `urlCache = nil` so neither response bodies nor metadata
+    /// touch disk.
+    private static func makeNonCachingSession() -> URLSession {
+        let config = URLSessionConfiguration.ephemeral
+        config.urlCache = nil
+        config.requestCachePolicy = .reloadIgnoringLocalCacheData
+        return URLSession(configuration: config)
+    }
+
    func startPolling(interval: TimeInterval = 0.5) {
        stopPolling()
        Task {
@@ -0,0 +1,242 @@
+import AppKit
+import SwiftUI
+
+/// Manages a standalone window for the bug-report flow.
+/// Ensures only one instance exists and brings it to front on repeated opens.
+@MainActor
+final class BugReportWindowController: ObservableObject {
+    private var window: NSWindow?
+
+    func open() {
+        if let existing = window, existing.isVisible {
+            existing.makeKeyAndOrderFront(nil)
+            NSApp.activate()
+            return
+        }
+
+        let view = BugReportView(onDismiss: { [weak self] in
+            self?.window?.close()
+        })
+
+        let hostingController = NSHostingController(rootView: view)
+        hostingController.sizingOptions = [.preferredContentSize, .minSize]
+
+        let newWindow = NSWindow(contentViewController: hostingController)
+        newWindow.styleMask = [.titled, .closable, .resizable]
+        newWindow.title = "Send a Bug Report"
+        newWindow.center()
+        newWindow.setFrameAutosaveName("ExoBugReportWindow")
+        newWindow.isReleasedWhenClosed = false
+        newWindow.makeKeyAndOrderFront(nil)
+        NSApp.activate()
+
+        window = newWindow
+    }
+}
+
+private struct BugReportView: View {
+    fileprivate enum Phase: Equatable {
+        case prompting
+        case sending(String)
+        case success(String)
+        case failure(String)
+    }
+
+    let onDismiss: () -> Void
+
+    @State private var phase: Phase = .prompting
+    @State private var userDescription: String = ""
+    @FocusState private var descriptionFocused: Bool
+
+    var body: some View {
+        VStack(alignment: .leading, spacing: 12) {
+            switch phase {
+            case .prompting:
+                promptingView
+            case .sending(let message):
+                sendingView(message: message)
+            case .success(let message):
+                successView(message: message)
+            case .failure(let message):
+                failureView(message: message)
+            }
+        }
+        .padding(16)
+        .frame(minWidth: 380)
+        .animation(.easeInOut(duration: 0.2), value: phase)
+        .onAppear { descriptionFocused = true }
+    }
+
+    private var promptingView: some View {
+        VStack(alignment: .leading, spacing: 8) {
+            Text("Description (optional)")
+                .font(.subheadline)
+                .foregroundColor(.secondary)
+            ZStack(alignment: .topLeading) {
+                if userDescription.isEmpty {
+                    Text("What were you doing when it broke?")
+                        .font(.body)
+                        .foregroundColor(Color(nsColor: .placeholderTextColor))
+                        .padding(.horizontal, 10)
+                        .padding(.vertical, 8)
+                        .allowsHitTesting(false)
+                }
+                TextEditor(text: $userDescription)
+                    .font(.body)
+                    .scrollContentBackground(.hidden)
+                    .padding(4)
+                    .frame(height: 72)
+                    .focused($descriptionFocused)
+            }
+            .background(
+                RoundedRectangle(cornerRadius: 6)
+                    .fill(Color(nsColor: .textBackgroundColor))
+            )
+            .overlay(
+                RoundedRectangle(cornerRadius: 6)
+                    .strokeBorder(Color(nsColor: .separatorColor), lineWidth: 1)
+            )
+
+            Text("Diagnostic logs will be uploaded with your report.")
+                .font(.caption)
+                .foregroundColor(.secondary)
+
+            HStack {
+                Spacer()
+                Button("Cancel") { onDismiss() }
+                    .keyboardShortcut(.cancelAction)
+                Button("Send") {
+                    Task { await send() }
+                }
+                .keyboardShortcut(.defaultAction)
+            }
+            .padding(.top, 4)
+        }
+    }
+
+    private func sendingView(message: String) -> some View {
+        VStack(alignment: .leading, spacing: 12) {
+            HStack(spacing: 10) {
+                ProgressView().controlSize(.small)
+                Text(message)
+                    .foregroundColor(.secondary)
+            }
+            HStack {
+                Spacer()
+                Button("Cancel") { onDismiss() }
+                    .keyboardShortcut(.cancelAction)
+                    .disabled(true)
+                Button("Send") {}
+                    .disabled(true)
+            }
+        }
+    }
+
+    private func successView(message: String) -> some View {
+        VStack(alignment: .leading, spacing: 12) {
+            HStack(alignment: .top, spacing: 10) {
+                Image(systemName: "checkmark.circle.fill")
+                    .foregroundColor(.green)
+                    .font(.title2)
+                Text(message)
+                    .fixedSize(horizontal: false, vertical: true)
+            }
+            HStack {
+                Button {
+                    openGitHubIssue()
+                } label: {
+                    HStack(spacing: 4) {
+                        Image(systemName: "arrow.up.right.square")
+                        Text("Open GitHub Issue")
+                    }
+                }
+                Spacer()
+                Button("Done") { onDismiss() }
+                    .keyboardShortcut(.defaultAction)
+            }
+        }
+    }
+
+    private func failureView(message: String) -> some View {
+        VStack(alignment: .leading, spacing: 12) {
+            HStack(alignment: .top, spacing: 10) {
+                Image(systemName: "exclamationmark.triangle.fill")
+                    .foregroundColor(.orange)
+                    .font(.title2)
+                Text(message)
+                    .fixedSize(horizontal: false, vertical: true)
+            }
+            HStack {
+                Spacer()
+                Button("Try Again") {
+                    phase = .prompting
+                }
+                Button("Close") { onDismiss() }
+                    .keyboardShortcut(.defaultAction)
+            }
+        }
+    }
+
+    private func send() async {
+        phase = .sending("Collecting logs and uploading…")
+        let service = BugReportService()
+        let description = userDescription.trimmingCharacters(in: .whitespacesAndNewlines)
+        do {
+            let outcome = try await service.sendReport(
+                isManual: true,
+                userDescription: description.isEmpty ? nil : description
+            )
+            if outcome.success {
+                phase = .success(outcome.message)
+            } else {
+                phase = .failure(outcome.message)
+            }
+        } catch {
+            phase = .failure(error.localizedDescription)
+        }
+    }
+
+    private func openGitHubIssue() {
+        let description = userDescription.trimmingCharacters(in: .whitespacesAndNewlines)
+
+        var bodyParts: [String] = []
+        bodyParts.append("## Describe the bug")
+        bodyParts.append("")
+        if !description.isEmpty {
+            bodyParts.append(description)
+        } else {
+            bodyParts.append("A clear and concise description of what the bug is.")
+        }
+        bodyParts.append("")
+        bodyParts.append("## Environment")
+        bodyParts.append("")
+        bodyParts.append("- macOS Version: \(ProcessInfo.processInfo.operatingSystemVersionString)")
+        bodyParts.append("- EXO Version: \(buildTag) (\(buildCommit))")
+        bodyParts.append("")
+        bodyParts.append("## Additional context")
+        bodyParts.append("")
+        bodyParts.append("A bug report with diagnostic logs was submitted via the app.")
+
+        let body = bodyParts.joined(separator: "\n")
+
+        var components = URLComponents(string: "https://github.com/exo-explore/exo/issues/new")!
+        components.queryItems = [
+            URLQueryItem(name: "template", value: "bug_report.md"),
+            URLQueryItem(name: "title", value: "[BUG] "),
+            URLQueryItem(name: "body", value: body),
+            URLQueryItem(name: "labels", value: "bug"),
+        ]
+
+        if let url = components.url {
+            NSWorkspace.shared.open(url)
+        }
+    }
+
+    private var buildTag: String {
+        Bundle.main.infoDictionary?["EXOBuildTag"] as? String ?? "unknown"
+    }
+
+    private var buildCommit: String {
+        Bundle.main.infoDictionary?["EXOBuildCommit"] as? String ?? "unknown"
+    }
+}
@@ -15,9 +15,12 @@ struct SettingsView: View {
    @State private var pendingHFEndpoint: String = ""
    @State private var pendingEnableImageModels = false
    @State private var pendingOfflineMode = false
+    @State private var pendingFastSynchEnabled = false
+    @State private var pendingDefaultModelsDir: String = ""
+    @State private var pendingAdditionalModelsDirs: String = ""
+    @State private var pendingReadOnlyModelsDirs: String = ""
+    @State private var pendingCustomEnvironmentVariables: [CustomEnvironmentVariable] = []
    @State private var needsRestart = false
-    @State private var bugReportInFlight = false
-    @State private var bugReportMessage: String?
    @State private var uninstallInProgress = false

    var body: some View {
@@ -34,18 +37,27 @@ struct SettingsView: View {
                .tabItem {
                    Label("Advanced", systemImage: "wrench.and.screwdriver")
                }
+            environmentTab
+                .tabItem {
+                    Label("Environment", systemImage: "terminal")
+                }
            aboutTab
                .tabItem {
                    Label("About", systemImage: "info.circle")
                }
        }
-        .frame(width: 450, height: 400)
+        .frame(width: 640, height: 560)
        .onAppear {
            pendingNamespace = controller.customNamespace
            pendingHFToken = controller.hfToken
            pendingHFEndpoint = controller.hfEndpoint
            pendingEnableImageModels = controller.enableImageModels
            pendingOfflineMode = controller.offlineMode
+            pendingFastSynchEnabled = controller.fastSynchEnabled
+            pendingDefaultModelsDir = controller.defaultModelsDir
+            pendingAdditionalModelsDirs = controller.additionalModelsDirs
+            pendingReadOnlyModelsDirs = controller.readOnlyModelsDirs
+            pendingCustomEnvironmentVariables = controller.customEnvironmentVariables
            needsRestart = false
        }
    }
@@ -56,9 +68,9 @@ struct SettingsView: View {
        Form {
            Section {
                LabeledContent("Cluster Namespace") {
-                    TextField("default", text: $pendingNamespace)
+                    TextField("", text: $pendingNamespace, prompt: Text("default"))
                        .textFieldStyle(.roundedBorder)
-                        .frame(width: 200)
+                        .frame(width: 260)
                }
                Text("Nodes with the same namespace form a cluster. Leave empty for default.")
                    .font(.caption)
@@ -67,9 +79,9 @@ struct SettingsView: View {

            Section {
                LabeledContent("HuggingFace Token") {
-                    SecureField("optional", text: $pendingHFToken)
+                    SecureField("", text: $pendingHFToken, prompt: Text("optional"))
                        .textFieldStyle(.roundedBorder)
-                        .frame(width: 200)
+                        .frame(width: 260)
                }
                Text("Required for gated models. Get yours at huggingface.co/settings/tokens")
                    .font(.caption)
@@ -78,9 +90,9 @@ struct SettingsView: View {

            Section {
                LabeledContent("HuggingFace Endpoint") {
-                    TextField("default", text: $pendingHFEndpoint)
+                    TextField("", text: $pendingHFEndpoint, prompt: Text("default"))
                        .textFieldStyle(.roundedBorder)
-                        .frame(width: 200)
+                        .frame(width: 260)
                }
                Text("Defaults to huggingface.co. Use a mirror (e.g. hf-mirror.com) for China.")
                    .font(.caption)
@@ -137,6 +149,23 @@ struct SettingsView: View {

    private var advancedTab: some View {
        Form {
+            Section("Performance") {
+                Toggle("Fast Synch Enabled", isOn: $pendingFastSynchEnabled)
+                Text(
+                    "Experimental: enables fast CPU to GPU synchronization. Can sometimes cause a \"GPU lock\" where inference hangs for ~10 seconds before starting. Necessary for low latency with RDMA and Tensor Parallelism."
+                )
+                .font(.caption)
+                .foregroundColor(.secondary)
+
+                HStack {
+                    Spacer()
+                    Button("Save & Restart") {
+                        applyAdvancedSettings()
+                    }
+                    .disabled(!hasAdvancedChanges)
+                }
+            }
+
            Section("Onboarding") {
                HStack {
                    VStack(alignment: .leading) {
@@ -171,8 +200,6 @@ struct SettingsView: View {
                VStack(alignment: .leading, spacing: 2) {
                    rdmaStatusView
                }
-
-                sendBugReportButton
            }

            Section("Danger Zone") {
@@ -193,6 +220,128 @@ struct SettingsView: View {
        .padding()
    }

+    // MARK: - Environment Tab
+
+    private var environmentTab: some View {
+        Form {
+            Section("Models Directories") {
+                LabeledContent("Default Models Directory") {
+                    TextField(
+                        "",
+                        text: $pendingDefaultModelsDir,
+                        prompt: Text("~/.exo/models")
+                    )
+                    .textFieldStyle(.roundedBorder)
+                    .font(.system(.body, design: .monospaced))
+                    .frame(width: 260)
+                }
+                Text("Sets EXO_DEFAULT_MODELS_DIR. Where models are downloaded.")
+                    .font(.caption)
+                    .foregroundColor(.secondary)
+
+                LabeledContent("Additional Directories") {
+                    TextField(
+                        "",
+                        text: $pendingAdditionalModelsDirs,
+                        prompt: Text("optional, colon-separated")
+                    )
+                    .textFieldStyle(.roundedBorder)
+                    .font(.system(.body, design: .monospaced))
+                    .frame(width: 260)
+                }
+                Text("Sets EXO_MODELS_DIRS. Extra writable model directories.")
+                    .font(.caption)
+                    .foregroundColor(.secondary)
+
+                LabeledContent("Read-Only Directories") {
+                    TextField(
+                        "",
+                        text: $pendingReadOnlyModelsDirs,
+                        prompt: Text("optional, colon-separated")
+                    )
+                    .textFieldStyle(.roundedBorder)
+                    .font(.system(.body, design: .monospaced))
+                    .frame(width: 260)
+                }
+                Text("Sets EXO_MODELS_READ_ONLY_DIRS. Never written to.")
+                    .font(.caption)
+                    .foregroundColor(.secondary)
+            }
+
+            Section("Custom Environment Variables") {
+                Text(
+                    "Escape hatch for env vars that don't have typed fields above. "
+                        + "Values here override the typed fields on conflict."
+                )
+                .font(.caption)
+                .foregroundColor(.secondary)
+
+                if pendingCustomEnvironmentVariables.isEmpty {
+                    Text("No custom variables.")
+                        .font(.caption)
+                        .foregroundColor(.secondary)
+                } else {
+                    ForEach($pendingCustomEnvironmentVariables) { $variable in
+                        HStack(alignment: .center, spacing: 8) {
+                            VStack(spacing: 4) {
+                                TextField("key", text: $variable.key)
+                                    .labelsHidden()
+                                    .textFieldStyle(.roundedBorder)
+                                    .font(.system(.body, design: .monospaced))
+                                TextField("value", text: $variable.value)
+                                    .labelsHidden()
+                                    .textFieldStyle(.roundedBorder)
+                                    .font(.system(.body, design: .monospaced))
+                            }
+                            VStack(spacing: 4) {
+                                Button {
+                                    pendingCustomEnvironmentVariables.removeAll {
+                                        $0.id == variable.id
+                                    }
+                                } label: {
+                                    Image(systemName: "minus.circle")
+                                }
+                                .buttonStyle(.borderless)
+                                .help("Remove variable")
+                                if !isValidEnvironmentVariableName(variable.key) {
+                                    Image(systemName: "exclamationmark.triangle.fill")
+                                        .foregroundColor(.orange)
+                                        .help(
+                                            "Invalid environment variable name. "
+                                                + "Must match [A-Za-z_][A-Za-z0-9_]*."
+                                        )
+                                }
+                            }
+                        }
+                    }
+                }
+
+                HStack {
+                    Button {
+                        pendingCustomEnvironmentVariables.append(
+                            CustomEnvironmentVariable()
+                        )
+                    } label: {
+                        Label("Add Variable", systemImage: "plus")
+                    }
+                    Spacer()
+                }
+            }
+
+            Section {
+                HStack {
+                    Spacer()
+                    Button("Save & Restart") {
+                        applyEnvironmentSettings()
+                    }
+                    .disabled(!hasEnvironmentChanges)
+                }
+            }
+        }
+        .formStyle(.grouped)
+        .padding()
+    }
+
    // MARK: - About Tab

    private var aboutTab: some View {
@@ -351,63 +500,30 @@ struct SettingsView: View {
        }
    }

-    private var sendBugReportButton: some View {
-        VStack(alignment: .leading, spacing: 4) {
-            Button {
-                Task {
-                    await sendBugReport()
-                }
-            } label: {
-                HStack {
-                    if bugReportInFlight {
-                        ProgressView()
-                            .scaleEffect(0.6)
-                    }
-                    Text("Send Bug Report")
-                        .font(.caption)
-                        .fontWeight(.semibold)
-                    Spacer()
-                }
-            }
-            .disabled(bugReportInFlight)
-
-            if let message = bugReportMessage {
-                Text(message)
-                    .font(.caption2)
-                    .foregroundColor(.secondary)
-                    .fixedSize(horizontal: false, vertical: true)
-            }
-        }
-    }
-
    // MARK: - Actions

-    private func sendBugReport() async {
-        bugReportInFlight = true
-        bugReportMessage = "Collecting logs..."
-        let service = BugReportService()
-        do {
-            let outcome = try await service.sendReport(isManual: true)
-            bugReportMessage = outcome.message
-        } catch {
-            bugReportMessage = error.localizedDescription
-        }
-        bugReportInFlight = false
-    }
-
    private func showUninstallConfirmationAlert() {
        let alert = NSAlert()
        alert.messageText = "Uninstall EXO"
        alert.informativeText = """
-            This will remove EXO and all its system components:
+            This will remove EXO and all its components:

            • Network configuration daemon
            • Launch at login registration
            • EXO network location
+            • EXO data directory (~/.exo)

            The app will be moved to Trash.
            """
        alert.alertStyle = .warning
+
+        let checkbox = NSButton(
+            checkboxWithTitle: "Keep downloaded models (~/.exo/models)",
+            target: nil, action: nil)
+        checkbox.state = .off
+        checkbox.sizeToFit()
+        alert.accessoryView = checkbox
+
        alert.addButton(withTitle: "Uninstall")
        alert.addButton(withTitle: "Cancel")

@@ -417,11 +533,11 @@ struct SettingsView: View {

        let response = alert.runModal()
        if response == .alertFirstButtonReturn {
-            performUninstall()
+            performUninstall(keepModels: checkbox.state == .on)
        }
    }

-    private func performUninstall() {
+    private func performUninstall(keepModels: Bool) {
        uninstallInProgress = true

        controller.cancelPendingLaunch()
@@ -431,6 +547,7 @@ struct SettingsView: View {
        DispatchQueue.global(qos: .utility).async {
            do {
                try NetworkSetupHelper.uninstall()
+                try Self.removeExoDirectory(keepModels: keepModels)

                DispatchQueue.main.async {
                    LaunchAtLoginHelper.disable()
@@ -454,6 +571,23 @@ struct SettingsView: View {
        }
    }

+    private static func removeExoDirectory(keepModels: Bool) throws {
+        let fm = FileManager.default
+        let exoDir = ExoProcessController.exoDirectoryURL
+        guard fm.fileExists(atPath: exoDir.path) else { return }
+
+        if !keepModels {
+            try fm.removeItem(at: exoDir)
+            return
+        }
+
+        let contents = try fm.contentsOfDirectory(
+            at: exoDir, includingPropertiesForKeys: nil, options: [])
+        for entry in contents where entry.lastPathComponent != "models" {
+            try? fm.removeItem(at: entry)
+        }
+    }
+
    private func moveAppToTrash() {
        guard let appURL = Bundle.main.bundleURL as URL? else { return }
        do {
@@ -475,6 +609,17 @@ struct SettingsView: View {
        pendingEnableImageModels != controller.enableImageModels
    }

+    private var hasAdvancedChanges: Bool {
+        pendingFastSynchEnabled != controller.fastSynchEnabled
+    }
+
+    private var hasEnvironmentChanges: Bool {
+        pendingDefaultModelsDir != controller.defaultModelsDir
+            || pendingAdditionalModelsDirs != controller.additionalModelsDirs
+            || pendingReadOnlyModelsDirs != controller.readOnlyModelsDirs
+            || pendingCustomEnvironmentVariables != controller.customEnvironmentVariables
+    }
+
    private func applyGeneralSettings() {
        controller.customNamespace = pendingNamespace
        controller.hfToken = pendingHFToken
@@ -488,6 +633,75 @@ struct SettingsView: View {
        restartIfRunning()
    }

+    private func applyAdvancedSettings() {
+        controller.fastSynchEnabled = pendingFastSynchEnabled
+        restartIfRunning()
+    }
+
+    private func applyEnvironmentSettings() {
+        controller.defaultModelsDir = pendingDefaultModelsDir.trimmingCharacters(
+            in: .whitespaces)
+        controller.additionalModelsDirs = pendingAdditionalModelsDirs.trimmingCharacters(
+            in: .whitespaces)
+        controller.readOnlyModelsDirs = pendingReadOnlyModelsDirs.trimmingCharacters(
+            in: .whitespaces)
+
+        pendingDefaultModelsDir = controller.defaultModelsDir
+        pendingAdditionalModelsDirs = controller.additionalModelsDirs
+        pendingReadOnlyModelsDirs = controller.readOnlyModelsDirs
+
+        // Trim whitespace from keys and drop empty ones so that the stored
+        // form matches what is actually injected into the child process and
+        // hasEnvironmentChanges doesn't show a stale diff after save.
+        let trimmed: [CustomEnvironmentVariable] =
+            pendingCustomEnvironmentVariables.compactMap { variable in
+                let key = variable.key.trimmingCharacters(in: .whitespaces)
+                guard !key.isEmpty else { return nil }
+                return CustomEnvironmentVariable(
+                    id: variable.id, key: key, value: variable.value
+                )
+            }
+
+        // De-duplicate keys, keeping the last occurrence. This matches the
+        // effective semantics of the dictionary assignment in
+        // ExoProcessController.makeEnvironment and avoids silently losing
+        // visible rows after save.
+        var seenKeys = Set<String>()
+        var deduplicatedReversed: [CustomEnvironmentVariable] = []
+        for variable in trimmed.reversed() {
+            if seenKeys.insert(variable.key).inserted {
+                deduplicatedReversed.append(variable)
+            }
+        }
+        let sanitized = Array(deduplicatedReversed.reversed())
+
+        pendingCustomEnvironmentVariables = sanitized
+        controller.customEnvironmentVariables = sanitized
+
+        restartIfRunning()
+    }
+
+    /// Validates a POSIX-style environment variable name:
+    /// `[A-Za-z_][A-Za-z0-9_]*`. Uses an ASCII-only charset so that
+    /// Unicode letters (e.g. `ñ`, Cyrillic) are rejected in line with what
+    /// the help tooltip advertises. Empty strings are treated as valid
+    /// here so that a freshly added blank row does not immediately look
+    /// broken; the save step filters empty keys out instead.
+    private func isValidEnvironmentVariableName(_ key: String) -> Bool {
+        if key.isEmpty { return true }
+        let headAllowed = CharacterSet(
+            charactersIn: "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz_"
+        )
+        let tailAllowed = headAllowed.union(CharacterSet(charactersIn: "0123456789"))
+        guard let first = key.unicodeScalars.first, headAllowed.contains(first) else {
+            return false
+        }
+        for scalar in key.unicodeScalars.dropFirst() {
+            if !tailAllowed.contains(scalar) { return false }
+        }
+        return true
+    }
+
    private func restartIfRunning() {
        if controller.status == .running || controller.status == .starting {
            controller.restart()
@@ -30,7 +30,7 @@ final class SettingsWindowController: ObservableObject {
        let hostingView = NSHostingView(rootView: settingsView)

        let newWindow = NSWindow(
-            contentRect: NSRect(x: 0, y: 0, width: 450, height: 400),
+            contentRect: NSRect(x: 0, y: 0, width: 640, height: 560),
            styleMask: [.titled, .closable],
            backing: .buffered,
            defer: false
@@ -3,25 +3,55 @@
 # EXO Uninstaller Script
 #
 # This script removes all EXO system components that persist after deleting the app.
-# Run with: sudo ./uninstall-exo.sh
+# Run with: sudo ./uninstall-exo.sh [--keep-models]
+#
+# Options:
+#   --keep-models   Preserve ~/.exo/models when removing the EXO data directory.
 #
 # Components removed:
 # - LaunchDaemon: /Library/LaunchDaemons/io.exo.networksetup.plist
 # - Network script: /Library/Application Support/EXO/
 # - Log files: /var/log/io.exo.networksetup.*
 # - Network location: "exo"
+# - EXO data directory: ~/.exo (or all of ~/.exo except models/ when --keep-models is set)
 # - Launch at login registration
 #

 set -euo pipefail

+KEEP_MODELS=0
+for arg in "$@"; do
+  case "$arg" in
+  --keep-models)
+    KEEP_MODELS=1
+    ;;
+  -h | --help)
+    echo "Usage: sudo ./uninstall-exo.sh [--keep-models]"
+    echo "  --keep-models   Preserve ~/.exo/models when removing the EXO data directory."
+    exit 0
+    ;;
+  *)
+    echo "Unknown argument: $arg" >&2
+    echo "Usage: sudo ./uninstall-exo.sh [--keep-models]" >&2
+    exit 2
+    ;;
+  esac
+done
+
 LABEL="io.exo.networksetup"
-SCRIPT_DEST="/Library/Application Support/EXO/disable_bridge_enable_dhcp.sh"
+# Current script path. Older installs used a different filename; keep the
+# legacy path here so a fresh uninstall still cleans up upgraded machines.
+CURRENT_SCRIPT_DEST="/Library/Application Support/EXO/disable_bridge.sh"
+LEGACY_SCRIPT_DEST="/Library/Application Support/EXO/disable_bridge_enable_dhcp.sh"
 PLIST_DEST="/Library/LaunchDaemons/io.exo.networksetup.plist"
 LOG_OUT="/var/log/${LABEL}.log"
 LOG_ERR="/var/log/${LABEL}.err.log"
 APP_BUNDLE_ID="io.exo.EXO"

+# Resolve the invoking user's home, even when run via sudo.
+USER_HOME="$(eval echo "~${SUDO_USER:-$USER}")"
+EXO_DIR="$USER_HOME/.exo"
+
 # Colors for output
 RED='\033[0;31m'
 GREEN='\033[0;32m'
@@ -69,11 +99,17 @@ else
  echo_warn "LaunchDaemon plist not found (already removed?)"
 fi

-# Remove the script and parent directory
-if [[ -f $SCRIPT_DEST ]]; then
-  rm -f "$SCRIPT_DEST"
-  echo_info "Removed network setup script"
-else
+# Remove the script (current and legacy filenames) — backwards-compatible:
+# tolerate either, both, or neither being present.
+removed_any_script=0
+for script in "$CURRENT_SCRIPT_DEST" "$LEGACY_SCRIPT_DEST"; do
+  if [[ -f $script ]]; then
+    rm -f "$script"
+    echo_info "Removed network setup script: $script"
+    removed_any_script=1
+  fi
+done
+if [[ $removed_any_script -eq 0 ]]; then
  echo_warn "Network setup script not found (already removed?)"
 fi

@@ -115,6 +151,22 @@ if networksetup -listnetworkservices 2>/dev/null | grep -q "Thunderbolt Bridge";
  echo_info "Re-enabled Thunderbolt Bridge"
 fi

+# Remove EXO data directory (~/.exo)
+EXO_DIR_REMOVED=""
+if [[ -d $EXO_DIR ]]; then
+  if [[ $KEEP_MODELS == "1" && -d "$EXO_DIR/models" ]]; then
+    find "$EXO_DIR" -mindepth 1 -maxdepth 1 ! -name models -exec rm -rf {} +
+    EXO_DIR_REMOVED="kept_models"
+    echo_info "Removed ~/.exo (preserved models/)"
+  else
+    rm -rf "$EXO_DIR"
+    EXO_DIR_REMOVED="full"
+    echo_info "Removed ~/.exo"
+  fi
+else
+  echo_warn "~/.exo not found (already removed?)"
+fi
+
 # Note about launch at login registration
 # SMAppService-based login items cannot be removed from a shell script.
 # They can only be unregistered from within the app itself or manually via System Settings.
@@ -144,6 +196,10 @@ echo "  • Network setup LaunchDaemon"
 echo "  • Network configuration script"
 echo "  • Log files"
 echo "  • 'exo' network location"
+case "$EXO_DIR_REMOVED" in
+full) echo "  • EXO data directory (~/.exo)" ;;
+kept_models) echo "  • EXO data directory (~/.exo, models preserved)" ;;
+esac
 echo ""
 echo "Your network has been restored to use the 'Automatic' location."
 echo "Thunderbolt Bridge has been re-enabled (if present)."
@@ -0,0 +1,160 @@
+# Exo-Bench — Methodology
+
+exo bench measures inference throughput and resource consumption of an exo cluster under controlled conditions. It sends prompts to the `/bench/chat/completions` endpoint, collects server-reported timing statistics, and records system-level metrics (power, GPU utilisation, temperature) throughout each run.
+
+The goal is to have accurate, transparent and reproducible numbers to compare speed and scaling across different models and different setups, and to be able to track these results as optimizations and features are added to EXO.
+
+Below is a technical summary of how Exo-Bench works. While the methodology and benchmark may change over time, this document will be kept up to date whenever this happens. If you find an issue with the methodology, or would like a  feature to be added, please open a GitHub issue!
+
+---
+
+## Prompt Construction
+
+Benchmarks need prompts of an exact token length. Unfortunately, we do not have direct access to the model but just the chat completion endpoint. To get around this fact, we create a request that will tokenise to a certain prompt length.
+
+This is achieved by:
+
+1. Tokenising a sample message through the model's `apply_chat_template()` to measure overhead (system tokens, special tokens, chat formatting).
+2. Binary-searching over a repeated atom string (default `"a "`) to find the content length that produces exactly the target number of tokens after template expansion.
+3. Returning both the content string and the verified token count.
+
+The actual token count is recorded in every result row as `pp_tokens`, so downstream analysis can confirm the prompt hit its target.
+
+Chat template formatting means that it may be impossible to attain very small pp benchmarks. e.g. pp=32 may not work. This tradeoff was made because the result of such a small prompt does not seem very interesting or useful for any real-world use cases.
+
+---
+
+## Bench Endpoint
+
+When a request reaches the server via the `/bench/chat/completions` endpoint, three things change compared to a normal chat completion:
+
+- **KV prefix cache is disabled by default**. Every request starts from a cold cache, ensuring prefill timing is not affected by prior requests. See [Prefix Cache Mode](#prefix-cache-mode) for the `--use-prefix-cache` option.
+- **EOS tokens are banned**. A logits processor suppresses all end-of-sequence tokens, forcing the model to generate exactly `max_tokens` tokens. This guarantees consistent generation length for fair TPS comparison — the model cannot short-circuit a run by stopping early.
+- **No model output parsing**. The bench collection path concatenates raw token text without any model-specific post-processing (thinking tag extraction, structured output handling, etc.). This is to avoid model outputs such as tool parsing or any structural mistakes from breaking the benchmark - we are testing for speed; see Exo-Eval for performance metrics.
+
+---
+
+## Timing
+
+### Prefill TPS
+
+Measured server-side per task.
+
+```
+prefill_tps = num_prompt_tokens / prefill_wall_seconds
+```
+
+### Generation TPS
+
+Measured server-side per task. Each task records wall-clock timestamps as tokens arrive:
+
+- First generated token: timestamp recorded
+- Every subsequent token: timestamp updated
+
+When generation completes:
+
+```
+gen_span = last_token_time - first_token_time
+generation_tps = (completion_tokens - 1) / gen_span
+```
+
+The first token is excluded from the numerator because the rate measures inter-token throughput — the time between the first and last token divided by the number of intervals.
+
+This does mean that tg=1 will not work.
+
+---
+
+## Concurrency
+
+### Single Request
+
+The client records wall-clock `elapsed_s` around the HTTP round-trip (network latency + server prefill + generation + response serialisation). This is a convenience metric for end-to-end latency. The authoritative TPS numbers come from the server-side per-task timing in the `generation_stats` response.
+
+### Concurrent Requests
+
+When `--concurrency N` is set with N > 1, all N requests must hit the server at the same instant. The mechanism:
+
+1. The prompt is built once and shared across all threads.
+2. Each thread gets its own HTTP connection.
+3. A thread barrier blocks all threads until every thread is ready.
+4. The first thread past the barrier records the batch start time and signals the others.
+5. All threads use the same start time as their reference, then fire their HTTP request.
+6. Each thread's `elapsed_s` is measured from the shared start time to its own response completion.
+
+**Batch wall time** is the maximum `elapsed_s` across all N requests — the time until the last request finishes.
+
+### Aggregate TPS
+
+```
+per_req_tps = max(generation_tps across N concurrent requests)
+agg_gen_tps = per_req_tps * concurrency
+```
+
+`max` is used instead of `mean` because all requests run in parallel against the same model. The fastest request's generation rate represents the system's per-stream throughput capacity; multiplying by concurrency gives aggregate throughput.
+
+---
+
+## Prefix Cache Mode
+
+When `--use-prefix-cache` is passed, the KV prefix cache remains active during benchmarking. This speeds up repeated runs by skipping redundant prefill work, which is useful when prompt processing is not the focus of the benchmark (e.g. when measuring generation throughput or power consumption across many configurations).
+
+Each response includes a `prefix_cache_hit` field (`"none"`, `"partial"`, or `"exact"`):
+
+- **none**: Cold prefill — no cached KV state was available. The reported `prompt_tps` is the real prefill throughput.
+- **partial**: A prefix of the prompt was found in cache. Only the remaining tokens were prefilled. The reported `prompt_tps` reflects the real throughput on the uncached portion. This occurs when multiple ascending `--pp` values share a common prefix (e.g. `--pp 1000,5000` — the 5000-token prompt reuses the 1000-token cache entry and prefills the remaining 4000 tokens).
+- **exact**: The entire prompt was found in cache (e.g. same `--pp` value on a `--repeat`). No prefill work was done. The reported `prompt_tps` is the TPS from when the cache entry was originally created, not a new measurement.
+
+**Prompt TPS is approximate in this mode.** Exact-hit runs report the stored TPS from the original cold/partial prefill rather than a freshly measured value. For accurate cold prefill numbers, run without `--use-prefix-cache`.
+
+Ascending `--pp` order (e.g. `--pp 1000,5000,10000`) gives the most useful data: each size gets a meaningful partial hit except the first which is cold. Descending order produces exact hits with approximate TPS from a longer prompt's original run.
+
+---
+
+## Warmup
+
+Before measurement begins, `--warmup N` (default: 0) discarded requests are sent using the first pp/tg pair. Warmup results are not included in the output.
+
+---
+
+## System Metrics
+
+A background thread polls each node at 1 Hz, collecting:
+
+- GPU utilisation (%)
+- Temperature (C)
+- System power draw (W)
+- CPU cluster usage (performance and efficiency cores)
+
+**Energy** is computed via trapezoidal integration of the power samples over each inference window (the wall-clock span of each benchmark request or concurrent batch). Average power is `total_joules / total_inference_seconds`.
+
+---
+
+## Output Format
+
+Results are written as JSON with three top-level keys:
+
+- **`runs`**: Array of per-request result objects, each containing:
+  - `elapsed_s`, `output_text_preview` (first 200 chars)
+  - `stats`: `{ prompt_tps, generation_tps, prompt_tokens, generation_tokens, peak_memory_usage }`
+  - Placement metadata: `model_id`, `placement_sharding`, `placement_instance_meta`, `placement_nodes`
+  - Run metadata: `pp_tokens`, `tg`, `repeat_index`, `concurrency`, `concurrent_index`
+  - `download_duration_s` (if model was freshly downloaded)
+- **`cluster`**: Cluster state snapshot at time of benchmarking.
+- **`system_metrics`**: Per-node time-series samples (GPU, power, temperature).
+
+---
+
+## Reproducing Results
+
+```bash
+cd bench && uv run python exo_bench.py \
+  --model "mlx-community/Qwen3.5-27B-4bit" \
+  --instance-meta jaccl \
+  --sharding tensor \
+  --min-nodes 2 --max-nodes 2 \
+  --pp 512 4096 --tg 128 \
+  --repeat 3 \
+  --warmup 1
+```
+
+Run --help for all the available flags.
@@ -7,7 +7,7 @@
 #   name, patterns, reasoning
 #
 # Optional per-model overrides (CLI flags take priority over these):
-#   temperature, top_p, max_tokens, reasoning_effort
+#   temperature, top_p, max_tokens, reasoning_effort, enable_thinking
 #
 # Fallback defaults (when no per-model config):
 #   reasoning:     temperature=1.0, max_tokens=131072, reasoning_effort="high"
@@ -18,10 +18,9 @@

 # ─── Qwen3.5 (Feb 2026) ─────────────────────────────────────────────
 # Source: HuggingFace model cards (Qwen/Qwen3.5-*)
-# 35B-A3B thinking general: temp=1.0, top_p=0.95, top_k=20
-# 397B thinking: temp=0.6, top_p=0.95, top_k=20
-# Non-thinking: temp=0.7, top_p=0.8, top_k=20
-# max_tokens: 32768 general, 81920 for complex math/code
+# Model card recommends: temp=0.6, top_p=0.95, top_k=20
+# We omit top_k to match vllm eval (which doesn't set it).
+# max_tokens=121072 to match vllm eval (131072 context - 10000 safety margin).

 [[model]]
 name = "Qwen3.5 2B"
@@ -29,7 +28,8 @@ patterns = ["Qwen3.5-2B"]
 reasoning = true
 temperature = 0.6
 top_p = 0.95
-max_tokens = 81920
+enable_thinking = true
+max_tokens = 121072

 [[model]]
 name = "Qwen3.5 9B"
@@ -37,7 +37,8 @@ patterns = ["Qwen3.5-9B"]
 reasoning = true
 temperature = 0.6
 top_p = 0.95
-max_tokens = 81920
+enable_thinking = true
+max_tokens = 121072

 [[model]]
 name = "Qwen3.5 27B"
@@ -45,15 +46,17 @@ patterns = ["Qwen3.5-27B"]
 reasoning = true
 temperature = 0.6
 top_p = 0.95
-max_tokens = 81920
+enable_thinking = true
+max_tokens = 121072

 [[model]]
 name = "Qwen3.5 35B A3B"
 patterns = ["Qwen3.5-35B-A3B"]
 reasoning = true
-temperature = 1.0
+temperature = 0.6
 top_p = 0.95
-max_tokens = 81920
+enable_thinking = true
+max_tokens = 121072

 [[model]]
 name = "Qwen3.5 122B A10B"
@@ -61,7 +64,8 @@ patterns = ["Qwen3.5-122B-A10B"]
 reasoning = true
 temperature = 0.6
 top_p = 0.95
-max_tokens = 81920
+enable_thinking = true
+max_tokens = 121072

 [[model]]
 name = "Qwen3.5 397B A17B"
@@ -69,12 +73,14 @@ patterns = ["Qwen3.5-397B-A17B"]
 reasoning = true
 temperature = 0.6
 top_p = 0.95
-max_tokens = 81920
+enable_thinking = true
+max_tokens = 121072

 # ─── Qwen3 (Apr 2025) ───────────────────────────────────────────────
 # Source: HuggingFace model cards (Qwen/Qwen3-*)
-# Thinking: temp=0.6, top_p=0.95, top_k=20
-# Non-thinking: temp=0.7, top_p=0.8, top_k=20
+# Model card recommends: temp=0.6, top_p=0.95, top_k=20
+# We omit top_k to match vllm eval (which doesn't set it).
+# Non-thinking: temp=0.7, top_p=0.8
 # max_tokens: 32768 general, 38912 for complex math/code

 [[model]]
@@ -83,6 +89,7 @@ patterns = ["Qwen3-0.6B"]
 reasoning = true
 temperature = 0.6
 top_p = 0.95
+enable_thinking = true
 max_tokens = 38912

 [[model]]
@@ -91,6 +98,7 @@ patterns = ["Qwen3-30B-A3B"]
 reasoning = true
 temperature = 0.6
 top_p = 0.95
+enable_thinking = true
 max_tokens = 38912

 [[model]]
@@ -99,6 +107,7 @@ patterns = ["Qwen3-235B-A22B"]
 reasoning = true
 temperature = 0.6
 top_p = 0.95
+enable_thinking = true
 max_tokens = 38912

 [[model]]
@@ -107,6 +116,7 @@ patterns = ["Qwen3-Next-80B-A3B-Thinking"]
 reasoning = true
 temperature = 0.6
 top_p = 0.95
+enable_thinking = true
 max_tokens = 38912

 [[model]]
@@ -129,9 +139,9 @@ max_tokens = 16384
 name = "Qwen3 Coder Next"
 patterns = ["Qwen3-Coder-Next"]
 reasoning = false
-temperature = 0.7
-top_p = 0.8
-max_tokens = 16384
+temperature = 1.0
+top_p = 0.95
+max_tokens = 121072

 # ─── GPT-OSS (OpenAI) ───────────────────────────────────────────────
 # Source: OpenAI GitHub README + HuggingFace discussion #21
@@ -165,10 +175,38 @@ patterns = ["DeepSeek-V3.1"]
 reasoning = true
 temperature = 0.0

+[[model]]
+name = "DeepSeek V3.2"
+patterns = ["DeepSeek-V3.2"]
+reasoning = true
+temperature = 1.0
+top_p = 0.95
+enable_thinking = true
+
+# ─── NVIDIA Nemotron ───────────────────────────────────────────────────
+# Source: HuggingFace model cards
+# All variants: temp=1.0, top_p=0.95, enable_thinking=true
+
+[[model]]
+name = "Nemotron Cascade 2 30B A3B"
+patterns = ["Nemotron-Cascade-2-30B-A3B"]
+reasoning = true
+temperature = 1.0
+top_p = 0.95
+enable_thinking = true
+
+[[model]]
+name = "Nemotron 3 Super 120B A12B"
+patterns = ["Nemotron-3-Super-120B-A12B", "NVIDIA-Nemotron-3-Super-120B-A12B"]
+reasoning = true
+temperature = 1.0
+top_p = 0.95
+enable_thinking = true
+
 # ─── GLM (ZhipuAI / THUDM) ──────────────────────────────────────────
 # Source: HuggingFace model cards + generation_config.json + docs.z.ai
 # GLM 4.5+: temp=1.0, top_p=0.95
-# Reasoning tasks: 131072 max_tokens; coding/SWE tasks: temp=0.7
+# max_tokens=121072 to match vllm eval (131072 context - 10000 safety margin)

 [[model]]
 name = "GLM-5"
@@ -176,7 +214,8 @@ patterns = ["GLM-5"]
 reasoning = true
 temperature = 1.0
 top_p = 0.95
-max_tokens = 131072
+enable_thinking = true
+max_tokens = 121072

 [[model]]
 name = "GLM 4.5 Air"
@@ -191,7 +230,8 @@ patterns = ["GLM-4.7-"]
 reasoning = true
 temperature = 1.0
 top_p = 0.95
-max_tokens = 131072
+enable_thinking = true
+max_tokens = 121072
 # Note: matches both GLM-4.7 and GLM-4.7-Flash

 # ─── Kimi (Moonshot AI) ─────────────────────────────────────────────
@@ -213,7 +253,8 @@ patterns = ["Kimi-K2.5"]
 reasoning = true
 temperature = 1.0
 top_p = 0.95
-max_tokens = 131072
+enable_thinking = true
+max_tokens = 121072

 [[model]]
 name = "Kimi K2 Instruct"
@@ -223,7 +264,17 @@ temperature = 0.6

 # ─── MiniMax ─────────────────────────────────────────────────────────
 # Source: HuggingFace model cards + generation_config.json
-# All models: temp=1.0, top_p=0.95, top_k=40
+# All models: temp=1.0, top_p=0.95
+# max_tokens=90000 to match vllm eval (100000 context - 10000 safety margin)
+
+[[model]]
+name = "MiniMax M2.7"
+patterns = ["MiniMax-M2.7"]
+reasoning = true
+temperature = 1.0
+top_p = 0.95
+enable_thinking = true
+max_tokens = 90000

 [[model]]
 name = "MiniMax M2.5"
@@ -231,6 +282,8 @@ patterns = ["MiniMax-M2.5"]
 reasoning = true
 temperature = 1.0
 top_p = 0.95
+enable_thinking = true
+max_tokens = 90000

 [[model]]
 name = "MiniMax M2.1"
@@ -251,6 +304,8 @@ patterns = ["Step-3.5-Flash"]
 reasoning = true
 temperature = 1.0
 top_p = 0.95
+enable_thinking = true
+max_tokens = 121072

 # ─── Llama (Meta) ───────────────────────────────────────────────────
 # Source: generation_config.json + meta-llama/llama-models generation.py
@@ -3,11 +3,13 @@ from __future__ import annotations

 import argparse
 import contextlib
+import io
 import json
 import os
 import sys
 import time
 import tomllib
+from concurrent.futures import ThreadPoolExecutor, as_completed
 from dataclasses import dataclass, field
 from pathlib import Path
 from typing import Any, Literal
@@ -17,6 +19,7 @@ from harness import (
    ExoClient,
    ExoHttpError,
    add_common_instance_args,
+    capture_cluster_snapshot,
    instance_id_from_instance,
    nodes_used_in_instance,
    resolve_model_short_id,
@@ -208,7 +211,7 @@ def _openai_build_request(
        "model": model,
        "messages": messages,
        "tools": tools,
-        "max_tokens": 16384,
+        "max_tokens": 4096,
        "temperature": 0.0,
    }
    return "/v1/chat/completions", body
@@ -275,7 +278,7 @@ def _openai_build_followup(
        "model": model,
        "messages": followup_messages,
        "tools": tools,
-        "max_tokens": 16384,
+        "max_tokens": 4096,
        "temperature": 0.0,
    }
    return "/v1/chat/completions", body
@@ -378,7 +381,7 @@ def _claude_build_request(
        "model": model,
        "messages": claude_messages,
        "tools": claude_tools,
-        "max_tokens": 16384,
+        "max_tokens": 4096,
        "temperature": 0.0,
    }
    if system_content is not None:
@@ -488,7 +491,7 @@ def _claude_build_followup(
        "model": model,
        "messages": claude_messages,
        "tools": claude_tools,
-        "max_tokens": 16384,
+        "max_tokens": 4096,
        "temperature": 0.0,
    }
    if system_content is not None:
@@ -912,6 +915,12 @@ Examples:
        default=1,
        help="Repeat each scenario N times (default: 1)",
    )
+    parser.add_argument(
+        "--concurrency",
+        type=int,
+        default=1,
+        help="Run up to N scenarios in parallel against the same instance (default: 1)",
+    )
    parser.add_argument(
        "--scenarios",
        nargs="*",
@@ -934,6 +943,13 @@ Examples:
    )
    args = parser.parse_args()

+    if args.concurrency < 1:
+        print(
+            f"--concurrency must be >= 1 (got {args.concurrency})",
+            file=sys.stderr,
+        )
+        sys.exit(2)
+
    all_scenarios = load_scenarios(SCENARIOS_PATH)
    if args.scenarios:
        scenarios = [s for s in all_scenarios if s.name in args.scenarios]
@@ -1006,44 +1022,75 @@ Examples:
        sys.exit(1)

    time.sleep(1)
+    cluster_snapshot = capture_cluster_snapshot(exo)
    all_results: list[ScenarioResult] = []

+    tasks: list[tuple[int, Scenario, ApiName]] = [
+        (run_idx, scenario, api_name)
+        for run_idx in range(args.repeat)
+        for scenario in scenarios
+        for api_name in api_names
+    ]
+
+    def _run_one(
+        http_client: httpx.Client,
+        task: tuple[int, Scenario, ApiName],
+    ) -> tuple[tuple[int, Scenario, ApiName], list[ScenarioResult], str]:
+        run_idx, scenario, api_name = task
+        buf = io.StringIO()
+        run_tag = f"[run {run_idx + 1}/{args.repeat}]" if args.repeat > 1 else ""
+        print(
+            f"\n  {run_tag}[{api_name:>9}] {scenario.name}: {scenario.description}",
+            file=buf,
+        )
+        scenario_results = run_scenario(
+            http_client,
+            args.host,
+            args.port,
+            full_model_id,
+            scenario,
+            api_name,
+            args.timeout,
+            args.verbose,
+        )
+        for r in scenario_results:
+            status = "PASS" if r.passed else "FAIL"
+            print(
+                f"    [{r.phase:>10}] {status}  ({r.latency_ms:.0f}ms)",
+                file=buf,
+            )
+            for check_name, check_ok in r.checks.items():
+                mark = "+" if check_ok else "-"
+                print(f"      {mark} {check_name}", file=buf)
+            if r.error:
+                print(f"      ! {r.error}", file=buf)
+        return task, scenario_results, buf.getvalue()
+
    try:
        with httpx.Client() as http_client:
-            for run_idx in range(args.repeat):
-                if args.repeat > 1:
-                    print(f"\n--- Run {run_idx + 1}/{args.repeat} ---", file=log)
-
-                for scenario in scenarios:
-                    for api_name in api_names:
-                        print(
-                            f"\n  [{api_name:>9}] {scenario.name}: {scenario.description}",
-                            file=log,
-                        )
-
-                        scenario_results = run_scenario(
-                            http_client,
-                            args.host,
-                            args.port,
-                            full_model_id,
-                            scenario,
-                            api_name,
-                            args.timeout,
-                            args.verbose,
-                        )
+            if args.concurrency == 1:
+                current_run = -1
+                for task in tasks:
+                    run_idx = task[0]
+                    if args.repeat > 1 and run_idx != current_run:
+                        print(f"\n--- Run {run_idx + 1}/{args.repeat} ---", file=log)
+                        current_run = run_idx
+                    _, scenario_results, buffered = _run_one(http_client, task)
+                    all_results.extend(scenario_results)
+                    log.write(buffered)
+                    log.flush()
+            else:
+                print(
+                    f"Running {len(tasks)} tasks with concurrency={args.concurrency}",
+                    file=log,
+                )
+                with ThreadPoolExecutor(max_workers=args.concurrency) as pool:
+                    futures = [pool.submit(_run_one, http_client, t) for t in tasks]
+                    for fut in as_completed(futures):
+                        _, scenario_results, buffered = fut.result()
                        all_results.extend(scenario_results)
-
-                        for r in scenario_results:
-                            status = "PASS" if r.passed else "FAIL"
-                            print(
-                                f"    [{r.phase:>10}] {status}  ({r.latency_ms:.0f}ms)",
-                                file=log,
-                            )
-                            for check_name, check_ok in r.checks.items():
-                                mark = "+" if check_ok else "-"
-                                print(f"      {mark} {check_name}", file=log)
-                            if r.error:
-                                print(f"      ! {r.error}", file=log)
+                        log.write(buffered)
+                        log.flush()
    finally:
        try:
            exo.request_json("DELETE", f"/instance/{instance_id}")
@@ -1084,16 +1131,19 @@ Examples:
                print(f"  - {r.name} [{r.api}/{r.phase}]: {r.error}", file=log)

    json_results = [result_to_dict(r) for r in all_results]
+    output: dict[str, Any] = {"results": json_results}
+    if cluster_snapshot:
+        output["cluster"] = cluster_snapshot

    if args.stdout:
-        print(json.dumps(json_results, indent=2))
+        print(json.dumps(output, indent=2))
    else:
        json_path = args.json_out
        parent = os.path.dirname(json_path)
        if parent:
            os.makedirs(parent, exist_ok=True)
        with open(json_path, "w") as f:
-            json.dump(json_results, f, indent=2)
+            json.dump(output, f, indent=2)
            f.write("\n")
        print(f"\nJSON results written to {json_path}", file=log)

@@ -22,6 +22,7 @@ import contextlib
 import itertools
 import json
 import sys
+import threading
 import time
 from collections.abc import Callable
 from concurrent.futures import ThreadPoolExecutor, as_completed
@@ -33,7 +34,10 @@ from harness import (
    ExoClient,
    ExoHttpError,
    add_common_instance_args,
+    capture_cluster_snapshot,
+    find_existing_instance,
    instance_id_from_instance,
+    node_ids_from_instance,
    nodes_used_in_instance,
    resolve_model_short_id,
    run_planning_phase,
@@ -76,7 +80,7 @@ def load_tokenizer_for_bench(model_id: str) -> Any:
        model_path = Path(
            snapshot_download(
                model_id,
-                allow_patterns=["*.json", "*.py", "*.tiktoken", "*.model"],
+                allow_patterns=["*.json", "*.py", "*.tiktoken", "*.model", "*.jinja"],
            )
        )

@@ -119,8 +123,48 @@ def load_tokenizer_for_bench(model_id: str) -> Any:

        return hf_tokenizer

-    # Default: use AutoTokenizer
-    return AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
+    # TODO: Change back to using only transformers
+    try:
+        return AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
+    except (AttributeError, ValueError):
+        from huggingface_hub import snapshot_download
+        from transformers import PretrainedConfig
+
+        model_path = Path(
+            snapshot_download(
+                model_id,
+                allow_patterns=[
+                    "*.json",
+                    "*.py",
+                    "tokenizer.model",
+                    "*.tiktoken",
+                    "tiktoken.model",
+                    "*.txt",
+                    "*.jsonl",
+                    "*.jinja",
+                ],
+            )
+        )
+        stub_kwargs: dict[str, Any] = {}
+        config_file = model_path / "config.json"
+        if config_file.exists():
+            with open(config_file) as f:
+                raw = json.load(f)
+            for key in (
+                "model_type",
+                "max_position_embeddings",
+                "vocab_size",
+                "bos_token_id",
+                "eos_token_id",
+                "pad_token_id",
+            ):
+                if key in raw:
+                    stub_kwargs[key] = raw[key]
+        return AutoTokenizer.from_pretrained(
+            str(model_path),
+            config=PretrainedConfig(**stub_kwargs),
+            trust_remote_code=True,
+        )


 def format_peak_memory(b: float) -> str:
@@ -131,6 +175,91 @@ def format_peak_memory(b: float) -> str:
    raise ValueError("You're using petabytes of memory. Something went wrong...")


+_SAMPLER_METRICS = ("gpuUsage", "temp", "sysPower", "pcpuUsage", "ecpuUsage")
+
+
+class SystemMetricsSampler:
+    def __init__(self, client: ExoClient, node_ids: list[str], interval_s: float = 1.0):
+        self._client = client
+        self._node_ids = node_ids
+        self._interval_s = interval_s
+        self._samples: dict[str, list[tuple[float, dict[str, float]]]] = {
+            nid: [] for nid in node_ids
+        }
+        self._stop = threading.Event()
+        self._thread: threading.Thread | None = None
+
+    def start(self) -> None:
+        self._stop.clear()
+        self._thread = threading.Thread(target=self._poll_loop, daemon=True)
+        self._thread.start()
+
+    def stop(self) -> None:
+        self._stop.set()
+        if self._thread:
+            self._thread.join(timeout=5)
+
+    def _poll_loop(self) -> None:
+        while not self._stop.is_set():
+            t = time.monotonic()
+            for nid in self._node_ids:
+                try:
+                    data = self._client.get_node_system(nid)
+                    if data:
+                        self._samples[nid].append(
+                            (t, {k: data.get(k, 0.0) for k in _SAMPLER_METRICS})
+                        )
+                except Exception:
+                    pass
+            self._stop.wait(self._interval_s)
+
+    def energy_between(self, t0: float, t1: float) -> float:
+        total_joules = 0.0
+        for _nid, samples in self._samples.items():
+            window = [(t, s["sysPower"]) for t, s in samples if t0 <= t <= t1]
+            if len(window) >= 2:
+                for i in range(1, len(window)):
+                    dt = window[i][0] - window[i - 1][0]
+                    avg_power = (window[i][1] + window[i - 1][1]) / 2
+                    total_joules += avg_power * dt
+            elif len(window) == 1:
+                total_joules += window[0][1] * (t1 - t0)
+        return total_joules
+
+    def summarize(self) -> dict[str, dict[str, dict[str, float]]]:
+        result: dict[str, dict[str, dict[str, float]]] = {}
+        for nid, samples in self._samples.items():
+            if not samples:
+                continue
+            metrics: dict[str, dict[str, float]] = {}
+            for key in _SAMPLER_METRICS:
+                values = [s[key] for t, s in samples]
+                metrics[key] = {
+                    "min": round(min(values), 2),
+                    "max": round(max(values), 2),
+                    "mean": round(mean(values), 2),
+                    "samples": len(values),
+                }
+            result[nid] = metrics
+        return result
+
+    def print_summary(self, placement_label: str) -> None:
+        summary = self.summarize()
+        if not summary:
+            return
+        logger.info(f"--- System Metrics ({placement_label}) ---")
+        for nid, metrics in summary.items():
+            gpu = metrics.get("gpuUsage", {})
+            temp = metrics.get("temp", {})
+            power = metrics.get("sysPower", {})
+            logger.info(
+                f"  {nid}:  "
+                f"GPU {gpu.get('mean', 0) * 100:.0f}% avg ({gpu.get('min', 0) * 100:.0f}–{gpu.get('max', 0) * 100:.0f}%)  |  "
+                f"{temp.get('mean', 0):.1f}°C avg  |  "
+                f"{power.get('mean', 0):.1f}W avg"
+            )
+
+
 def parse_int_list(values: list[str]) -> list[int]:
    items: list[int] = []
    for v in values:
@@ -142,27 +271,79 @@ def parse_int_list(values: list[str]) -> list[int]:


 def run_one_completion(
-    client: ExoClient, model_id: str, pp_hint: int, tg: int, prompt_sizer: PromptSizer
+    client: ExoClient,
+    model_id: str,
+    pp_hint: int,
+    tg: int,
+    prompt_sizer: PromptSizer,
+    *,
+    use_prefix_cache: bool = False,
+    stream: bool = False,
 ) -> tuple[dict[str, Any], int]:
    content, pp_tokens = prompt_sizer.build(pp_hint)
    payload: dict[str, Any] = {
        "model": model_id,
        "messages": [{"role": "user", "content": content}],
-        "stream": False,
        "max_tokens": tg,
+        "logprobs": False,
+        "use_prefix_cache": use_prefix_cache,
    }

-    t0 = time.perf_counter()
-    out = client.post_bench_chat_completions(payload)
-    elapsed = time.perf_counter() - t0
+    if not stream:
+        payload["stream"] = False
+        t0 = time.perf_counter()
+        out = client.post_bench_chat_completions(payload)
+        elapsed = time.perf_counter() - t0

-    stats = out.get("generation_stats")
+        stats = out.get("generation_stats")
+        choices = out.get("choices") or [{}]
+        message = choices[0].get("message", {}) if choices else {}
+        content = message.get("content") or ""
+        preview = content[:200] if content else ""
+    else:
+        tokens = 0
+        first_token_time = None
+        t0 = time.perf_counter()
+        text_parts: list[str] = []
+        stats = None

-    # Extract preview, handling None content (common for thinking models)
-    choices = out.get("choices") or [{}]
-    message = choices[0].get("message", {}) if choices else {}
-    content = message.get("content") or ""
-    preview = content[:200] if content else ""
+        for raw_line in client.stream_bench_chat_completions(payload):
+            line = raw_line.strip()
+            if line.startswith(": generation_stats "):
+                with contextlib.suppress(json.JSONDecodeError):
+                    stats = json.loads(line[len(": generation_stats ") :])
+                continue
+            if not line.startswith("data: "):
+                continue
+            data = line[6:]
+            if data == "[DONE]":
+                break
+            try:
+                chunk = json.loads(data)
+                delta = chunk.get("choices", [{}])[0].get("delta", {})
+                if delta.get("content"):
+                    if first_token_time is None:
+                        first_token_time = time.perf_counter()
+                    tokens += 1
+                    text_parts.append(delta["content"])
+            except json.JSONDecodeError:
+                pass
+
+        elapsed = time.perf_counter() - t0
+        preview = "".join(text_parts)[:200]
+
+        if not stats:
+            ttft = (first_token_time - t0) if first_token_time else elapsed
+            gen_time = elapsed - ttft if tokens > 1 else elapsed
+            gen_tps = (tokens - 1) / gen_time if tokens > 1 and gen_time > 0 else 0.0
+            prompt_tps = pp_tokens / ttft if ttft > 0 else 0.0
+            stats = {
+                "prompt_tokens": pp_tokens,
+                "generation_tokens": tokens,
+                "prompt_tps": round(prompt_tps, 2),
+                "generation_tps": round(gen_tps, 2),
+                "peak_memory_usage": {"inBytes": 0},
+            }

    return {
        "elapsed_s": elapsed,
@@ -182,9 +363,19 @@ class PromptSizer:
    def _make_counter(tokenizer: Any) -> Callable[[str], int]:
        def count_fn(user_content: str) -> int:
            messages = [{"role": "user", "content": user_content}]
-            ids = tokenizer.apply_chat_template(
-                messages, tokenize=True, add_generation_prompt=True
-            )
+            try:
+                ids = tokenizer.apply_chat_template(
+                    messages, tokenize=True, add_generation_prompt=True
+                )
+            except ValueError:
+                # Models without a Jinja chat template (e.g. DeepSeek V4 which
+                # ships its own Python encoder). Use the exo-side V4 encoder.
+                from exo.worker.engines.mlx.deepseek_v4_encoding import (
+                    encode_messages as encode_v4,
+                )
+
+                prompt = encode_v4(messages, thinking_mode="thinking")
+                ids = tokenizer.encode(prompt, add_special_tokens=False)
            # Fix for transformers 5.x
            if hasattr(ids, "input_ids"):
                ids = ids.input_ids
@@ -279,6 +470,27 @@ def main() -> int:
        action="store_true",
        help="Force all pp×tg combinations (cartesian product) even when lists have equal length.",
    )
+    ap.add_argument(
+        "--stream",
+        action="store_true",
+        help="Use /bench/chat/completions with streaming SSE response (bench=True still applies: no EOS detection, no KV cache).",
+    )
+    ap.add_argument(
+        "--no-system-metrics",
+        action="store_true",
+        help="Disable GPU utilization, temperature, and power collection during inference.",
+    )
+    ap.add_argument(
+        "--metrics-interval",
+        type=float,
+        default=1.0,
+        help="System metrics polling interval in seconds (default: 1.0).",
+    )
+    ap.add_argument(
+        "--use-prefix-cache",
+        action="store_true",
+        help="Enable KV prefix cache during bench (default: disabled for cold-cache measurements).",
+    )
    args = ap.parse_args()

    pp_list = parse_int_list(args.pp)
@@ -294,6 +506,15 @@ def main() -> int:
        logger.error("--concurrency values must be >= 1")
        return 2

+    if args.use_prefix_cache:
+        logger.warning(
+            "--use-prefix-cache: prompt TPS will be approximate. See METHODOLOGY.md for details."
+        )
+        if pp_list != sorted(pp_list):
+            logger.warning(
+                "--pp values are not in ascending order: prompt TPS will be less accurate. Use ascending --pp for best results."
+            )
+
    # Log pairing mode
    use_combinations = args.all_combinations or len(pp_list) != len(tg_list)
    if use_combinations:
@@ -319,82 +540,146 @@ def main() -> int:
        logger.error("[exo-bench] tokenizer usable but prompt sizing failed")
        raise

-    selected = settle_and_fetch_placements(
-        client, full_model_id, args, settle_timeout=args.settle_timeout
-    )
+    # Optionally reuse a running instance for this model
+    reused_instance_id: str | None = None
+    if args.reuse_instance:
+        existing = find_existing_instance(client, full_model_id)
+        if existing:
+            reused_instance_id = existing
+            logger.info(f"Reusing existing instance {reused_instance_id}")
+        else:
+            logger.warning(
+                "--reuse-instance: no existing instance found, creating a new one"
+            )

-    if not selected:
-        logger.error("No valid placements matched your filters.")
-        return 1
-
-    selected.sort(
-        key=lambda p: (
-            str(p.get("instance_meta", "")),
-            str(p.get("sharding", "")),
-            -nodes_used_in_instance(p["instance"]),
-        ),
-        reverse=True,
-    )
-
-    logger.debug(f"exo-bench model: short_id={short_id} full_id={full_model_id}")
-    logger.info(f"placements: {len(selected)}")
-    for p in selected:
-        logger.info(
-            f"  - {p['sharding']} / {p['instance_meta']} / nodes={nodes_used_in_instance(p['instance'])}"
+    if reused_instance_id is not None:
+        # Use the existing instance directly — skip placement iteration
+        selected = []
+        download_duration_s = None
+    else:
+        selected = settle_and_fetch_placements(
+            client, full_model_id, args, settle_timeout=args.settle_timeout
        )

-    if args.dry_run:
-        return 0
+        if not selected:
+            logger.error("No valid placements matched your filters.")
+            return 1

-    settle_deadline = (
-        time.monotonic() + args.settle_timeout if args.settle_timeout > 0 else None
-    )
+        selected.sort(
+            key=lambda p: (
+                str(p.get("instance_meta", "")),
+                str(p.get("sharding", "")),
+                nodes_used_in_instance(p["instance"]),
+            ),
+            reverse=True,
+        )

-    logger.info("Planning phase: checking downloads...")
-    download_duration_s = run_planning_phase(
-        client,
-        full_model_id,
-        selected[0],
-        args.danger_delete_downloads,
-        args.timeout,
-        settle_deadline,
-    )
-    if download_duration_s is not None:
-        logger.info(f"Download: {download_duration_s:.1f}s (freshly downloaded)")
-    else:
-        logger.info("Download: model already cached")
+        logger.debug(f"exo-bench model: short_id={short_id} full_id={full_model_id}")
+        logger.info(f"placements: {len(selected)}")
+        for p in selected:
+            logger.info(
+                f"  - {p['sharding']} / {p['instance_meta']} / nodes={nodes_used_in_instance(p['instance'])}"
+            )

+        if args.dry_run:
+            return 0
+
+        settle_deadline = (
+            time.monotonic() + args.settle_timeout if args.settle_timeout > 0 else None
+        )
+
+        logger.info("Planning phase: checking downloads...")
+        download_duration_s = run_planning_phase(
+            client,
+            full_model_id,
+            selected[0],
+            args.danger_delete_downloads,
+            args.timeout,
+            settle_deadline,
+        )
+        if download_duration_s is not None:
+            logger.info(f"Download: {download_duration_s:.1f}s (freshly downloaded)")
+        else:
+            logger.info("Download: model already cached")
+
+    cluster_snapshot = capture_cluster_snapshot(client)
    all_rows: list[dict[str, Any]] = []
+    all_system_metrics: dict[str, dict[str, dict[str, float]]] = {}
+
+    # If reusing an existing instance, run a single benchmark pass against it
+    if reused_instance_id is not None:
+        selected = [None]

    for preview in selected:
-        instance = preview["instance"]
-        instance_id = instance_id_from_instance(instance)
+        created_instance = False
+        if preview is not None:
+            instance = preview["instance"]
+            instance_id = instance_id_from_instance(instance)

-        sharding = str(preview["sharding"])
-        instance_meta = str(preview["instance_meta"])
-        n_nodes = nodes_used_in_instance(instance)
+            sharding = str(preview["sharding"])
+            instance_meta = str(preview["instance_meta"])
+            n_nodes = nodes_used_in_instance(instance)

-        logger.info("=" * 80)
-        logger.info(
-            f"PLACEMENT: {sharding} / {instance_meta} / nodes={n_nodes} / instance_id={instance_id}"
-        )
+            logger.info("=" * 80)
+            logger.info(
+                f"PLACEMENT: {sharding} / {instance_meta} / nodes={n_nodes} / instance_id={instance_id}"
+            )

-        client.request_json("POST", "/instance", body={"instance": instance})
-        try:
-            wait_for_instance_ready(client, instance_id)
-        except (RuntimeError, TimeoutError) as e:
-            logger.error(f"Failed to initialize placement: {e}")
-            with contextlib.suppress(ExoHttpError):
-                client.request_json("DELETE", f"/instance/{instance_id}")
-            continue
+            # Delete any existing instances to free resources before placing
+            try:
+                state = client.request_json("GET", "/state")
+                for old_id in list(state.get("instances", {}).keys()):
+                    logger.info(f"Deleting stale instance {old_id}")
+                    with contextlib.suppress(ExoHttpError):
+                        client.request_json("DELETE", f"/instance/{old_id}")
+                if state.get("instances"):
+                    time.sleep(2)
+            except Exception as e:
+                logger.warning(f"Failed to clean up stale instances: {e}")

-        time.sleep(1)
+            client.request_json("POST", "/instance", body={"instance": instance})
+            try:
+                wait_for_instance_ready(client, instance_id)
+            except (RuntimeError, TimeoutError) as e:
+                logger.error(f"Failed to initialize placement: {e}")
+                with contextlib.suppress(ExoHttpError):
+                    client.request_json("DELETE", f"/instance/{instance_id}")
+                continue
+
+            time.sleep(1)
+            created_instance = True
+        else:
+            instance_id = reused_instance_id
+            sharding = "reused"
+            instance_meta = "reused"
+            n_nodes = 0
+            logger.info("=" * 80)
+            logger.info(f"Using existing instance {instance_id}")
+
+        sampler: SystemMetricsSampler | None = None
+        if not args.no_system_metrics and preview is not None:
+            nids = node_ids_from_instance(instance)
+            sampler = SystemMetricsSampler(
+                ExoClient(args.host, args.port, timeout_s=30),
+                nids,
+                interval_s=args.metrics_interval,
+            )
+            sampler.start()
+
+        def _do_one(c: ExoClient, pp: int, tg: int) -> tuple[dict[str, Any], int]:
+            return run_one_completion(
+                c,
+                full_model_id,
+                pp,
+                tg,
+                prompt_sizer,
+                use_prefix_cache=args.use_prefix_cache,
+                stream=args.stream,
+            )

        try:
            for i in range(args.warmup):
-                run_one_completion(
-                    client, full_model_id, pp_list[0], tg_list[0], prompt_sizer
-                )
+                _do_one(client, pp_list[0], tg_list[0])
                logger.debug(f"  warmup {i + 1}/{args.warmup} done")

            # If pp and tg lists have same length, run in tandem (zip)
@@ -408,15 +693,16 @@ def main() -> int:
                for concurrency in concurrency_list:
                    logger.info(f"--- pp={pp} tg={tg} concurrency={concurrency} ---")
                    runs: list[dict[str, Any]] = []
+                    inference_windows: list[tuple[float, float]] = []
                    for r in range(args.repeat):
                        time.sleep(3)

                        if concurrency <= 1:
                            # Sequential: single request
                            try:
-                                row, actual_pp_tokens = run_one_completion(
-                                    client, full_model_id, pp, tg, prompt_sizer
-                                )
+                                inf_t0 = time.monotonic()
+                                row, actual_pp_tokens = _do_one(client, pp, tg)
+                                inference_windows.append((inf_t0, time.monotonic()))
                            except Exception as e:
                                logger.error(e)
                                continue
@@ -443,20 +729,54 @@ def main() -> int:
                            all_rows.append(row)
                        else:
                            # Concurrent: fire N requests in parallel
-                            # Each thread gets its own ExoClient (separate HTTP connection)
+                            # Pre-build prompt once, barrier ensures simultaneous dispatch
+                            content, actual_pp = prompt_sizer.build(pp)
+                            pre_built_payload: dict[str, Any] = {
+                                "model": full_model_id,
+                                "messages": [{"role": "user", "content": content}],
+                                "stream": False,
+                                "max_tokens": tg,
+                                "logprobs": False,
+                                "use_prefix_cache": args.use_prefix_cache,
+                            }
+                            barrier = threading.Barrier(concurrency)
+                            batch_start = threading.Event()
+                            batch_t0: float = 0.0
                            batch_results: list[tuple[dict[str, Any], int]] = []
                            batch_errors = 0

                            def _run_concurrent(
-                                idx: int, *, _pp: int = pp, _tg: int = tg
+                                idx: int,
+                                _barrier: threading.Barrier = barrier,
+                                _batch_start: threading.Event = batch_start,
+                                _payload: dict[str, Any] = pre_built_payload,
+                                _actual_pp: int = actual_pp,
                            ) -> tuple[dict[str, Any], int]:
+                                nonlocal batch_t0
                                c = ExoClient(
                                    args.host, args.port, timeout_s=args.timeout
                                )
-                                return run_one_completion(
-                                    c, full_model_id, _pp, _tg, prompt_sizer
+                                if _barrier.wait() == 0:
+                                    batch_t0 = time.perf_counter()
+                                    _batch_start.set()
+                                else:
+                                    _batch_start.wait()
+                                t0 = batch_t0
+                                out = c.post_bench_chat_completions(_payload)
+                                elapsed = time.perf_counter() - t0
+                                stats = out.get("generation_stats")
+                                choices = out.get("choices") or [{}]
+                                message = (
+                                    choices[0].get("message", {}) if choices else {}
                                )
+                                text = message.get("content") or ""
+                                return {
+                                    "elapsed_s": elapsed,
+                                    "output_text_preview": text[:200],
+                                    "stats": stats,
+                                }, _actual_pp

+                            inf_t0 = time.monotonic()
                            with ThreadPoolExecutor(max_workers=concurrency) as pool:
                                futures = {
                                    pool.submit(_run_concurrent, i): i
@@ -468,6 +788,12 @@ def main() -> int:
                                    except Exception as e:
                                        logger.error(f"Concurrent request failed: {e}")
                                        batch_errors += 1
+                            batch_wall_s = (
+                                max(x["elapsed_s"] for x, _ in batch_results)
+                                if batch_results
+                                else time.perf_counter() - batch_t0
+                            )
+                            inference_windows.append((inf_t0, time.monotonic()))

                            for idx, (row, actual_pp_tokens) in enumerate(
                                batch_results
@@ -502,48 +828,80 @@ def main() -> int:
                                    if x["stats"]["generation_tps"] > 0
                                ]
                                per_req_tps = (
-                                    mean(valid_gen_tps) if valid_gen_tps else 0.0
+                                    max(valid_gen_tps) if valid_gen_tps else 0.0
                                )
                                agg_gen_tps = per_req_tps * concurrency
                                logger.info(
                                    f"[concurrent {concurrency}x]  "
                                    f"agg_gen_tps={agg_gen_tps:.2f}  "
                                    f"per_req_tps={per_req_tps:.2f}  "
+                                    f"wall_s={batch_wall_s:.2f}  "
                                    f"errors={batch_errors}"
                                )

                    if runs:
                        prompt_tps = mean(x["stats"]["prompt_tps"] for x in runs)
-                        per_req_tps = mean(x["stats"]["generation_tps"] for x in runs)
+                        valid_gen = [
+                            x["stats"]["generation_tps"]
+                            for x in runs
+                            if x["stats"]["generation_tps"] > 0
+                        ]
+                        per_req_tps = max(valid_gen) if valid_gen else 0.0
                        gen_tps = per_req_tps * concurrency
                        ptok = mean(x["stats"]["prompt_tokens"] for x in runs)
                        gtok = mean(x["stats"]["generation_tokens"] for x in runs)
-                        peak = mean(
-                            x["stats"]["peak_memory_usage"]["inBytes"] for x in runs
-                        )

-                        logger.info(
+                        def _peak_bytes(s: dict[str, Any]) -> float:
+                            pm = s["peak_memory_usage"]
+                            return pm.get("inBytes") or pm.get("in_bytes", 0)
+
+                        peak = mean(_peak_bytes(x["stats"]) for x in runs)
+                        summary = (
                            f"prompt_tps={prompt_tps:.2f} gen_tps={gen_tps:.2f}    "
                            f"prompt_tokens={ptok} gen_tokens={gtok}    "
-                            f"peak_memory={format_peak_memory(peak)}\n"
+                            f"peak_memory={format_peak_memory(peak)}"
                        )
+                        if sampler and inference_windows:
+                            joules = sum(
+                                sampler.energy_between(t0, t1)
+                                for t0, t1 in inference_windows
+                            )
+                            inf_seconds = sum(t1 - t0 for t0, t1 in inference_windows)
+                            avg_watts = joules / inf_seconds if inf_seconds > 0 else 0
+                            summary += f"    energy={joules:.1f}J ({avg_watts:.1f}W avg over {inf_seconds:.1f}s inference)"
+                        logger.info(f"{summary}\n")
                    time.sleep(2)
        finally:
-            try:
-                client.request_json("DELETE", f"/instance/{instance_id}")
-            except ExoHttpError as e:
-                if e.status != 404:
-                    raise
-            wait_for_instance_gone(client, instance_id)
-            logger.debug(f"Deleted instance {instance_id}")
+            if sampler:
+                sampler.stop()
+                placement_label = f"{sharding}/{instance_meta}/{n_nodes} nodes"
+                sampler.print_summary(placement_label)
+                placement_metrics = sampler.summarize()
+                if placement_metrics:
+                    all_system_metrics.update(placement_metrics)

-            time.sleep(5)
+            if created_instance and instance_id is not None:
+                try:
+                    client.request_json("DELETE", f"/instance/{instance_id}")
+                except ExoHttpError as e:
+                    if e.status != 404:
+                        raise
+                wait_for_instance_gone(client, instance_id)
+                logger.debug(f"Deleted instance {instance_id}")
+
+                time.sleep(5)
+
+    output: dict[str, Any] = {"runs": all_rows}
+    if cluster_snapshot:
+        output["cluster"] = cluster_snapshot
+    if all_system_metrics:
+        output["system_metrics"] = all_system_metrics

    if args.stdout:
-        json.dump(all_rows, sys.stdout, indent=2, ensure_ascii=False)
+        json.dump(output, sys.stdout, indent=2, ensure_ascii=False)
    elif args.json_out:
        with open(args.json_out, "w", encoding="utf-8") as f:
-            json.dump(all_rows, f, indent=2, ensure_ascii=False)
+            json.dump(output, f, indent=2, ensure_ascii=False)
        logger.debug(f"\nWrote results JSON: {args.json_out}")

    return 0
@@ -46,6 +46,8 @@ from harness import (
    ExoClient,
    ExoHttpError,
    add_common_instance_args,
+    capture_cluster_snapshot,
+    find_existing_instance,
    instance_id_from_instance,
    nodes_used_in_instance,
    resolve_model_short_id,
@@ -61,6 +63,15 @@ from loguru import logger
 # ---------------------------------------------------------------------------

 MAX_RETRIES = 30
+INSTANCE_HEALTH_CHECK_AFTER = (
+    3  # Check instance health after this many consecutive failures
+)
+
+
+class InstanceFailedError(RuntimeError):
+    """Raised when the exo instance is detected as failed/gone."""
+
+
 DEFAULT_MAX_TOKENS = 16_384
 REASONING_MAX_TOKENS = 131_072
 TEMPERATURE_NON_REASONING = 0.0
@@ -270,7 +281,7 @@ def run_humaneval_test(

@dataclass
 class QuestionResult:
-    question_id: int
+    question_id: int | str
    prompt: str
    response: str
    extracted_answer: str | None
@@ -280,7 +291,11 @@ class QuestionResult:
    prompt_tokens: int = 0
    completion_tokens: int = 0
    reasoning_tokens: int = 0
+    reasoning_content: str = ""
+    finish_reason: str = ""
    elapsed_s: float = 0.0
+    power_watts: float = 0.0
+    energy_joules: float = 0.0


@dataclass
@@ -516,6 +531,10 @@ class ApiResult:
    prompt_tokens: int
    completion_tokens: int
    reasoning_tokens: int
+    reasoning_content: str = ""
+    finish_reason: str = ""
+    power_watts: float = 0.0
+    energy_joules: float = 0.0


 async def _call_api(
@@ -529,6 +548,9 @@ async def _call_api(
    system_message: str | None = None,
    reasoning_effort: str | None = None,
    top_p: float | None = None,
+    top_k: int | None = None,
+    min_p: float | None = None,
+    enable_thinking: bool | None = None,
 ) -> ApiResult:
    messages = []
    if system_message:
@@ -545,6 +567,12 @@ async def _call_api(
        body["reasoning_effort"] = reasoning_effort
    if top_p is not None:
        body["top_p"] = top_p
+    if top_k is not None:
+        body["top_k"] = top_k
+    if min_p is not None:
+        body["min_p"] = min_p
+    if enable_thinking is not None:
+        body["enable_thinking"] = enable_thinking

    resp = await client.post(
        f"{base_url}/v1/chat/completions",
@@ -553,19 +581,40 @@ async def _call_api(
    )
    resp.raise_for_status()
    data = resp.json()
-    content = data["choices"][0]["message"]["content"]
-    if not content or not content.strip():
+    choice = data["choices"][0]
+    message = choice["message"]
+    content = message.get("content") or ""
+    reasoning_content = message.get("reasoning_content") or ""
+    finish_reason = choice.get("finish_reason") or ""
+
+    # For thinking models, empty content is expected when finish_reason is "length"
+    if not content.strip() and finish_reason != "length" and not reasoning_content:
        raise ValueError("Empty response from model")
    usage = data.get("usage", {})
    details = usage.get("completion_tokens_details", {})
+    power = data.get("power_usage") or {}
    return ApiResult(
        content=content,
        prompt_tokens=usage.get("prompt_tokens", 0),
        completion_tokens=usage.get("completion_tokens", 0),
        reasoning_tokens=details.get("reasoning_tokens", 0) if details else 0,
+        reasoning_content=reasoning_content,
+        finish_reason=finish_reason,
+        power_watts=power.get("total_avg_sys_power_watts", 0.0),
+        energy_joules=power.get("total_energy_joules", 0.0),
    )


+async def _check_instance_health(base_url: str) -> bool:
+    """Return True if the exo instance is still reachable."""
+    try:
+        async with httpx.AsyncClient() as c:
+            resp = await c.get(f"{base_url}/models", timeout=5.0)
+            return resp.status_code == 200
+    except Exception:
+        return False
+
+
 async def call_with_retries(
    client: httpx.AsyncClient,
    base_url: str,
@@ -577,8 +626,14 @@ async def call_with_retries(
    system_message: str | None = None,
    reasoning_effort: str | None = None,
    top_p: float | None = None,
+    top_k: int | None = None,
+    min_p: float | None = None,
+    enable_thinking: bool | None = None,
+    instance_failed: asyncio.Event | None = None,
 ) -> ApiResult | None:
    for attempt in range(MAX_RETRIES):
+        if instance_failed and instance_failed.is_set():
+            raise InstanceFailedError("Instance already marked as failed")
        try:
            return await _call_api(
                client,
@@ -591,8 +646,30 @@ async def call_with_retries(
                system_message,
                reasoning_effort,
                top_p,
+                top_k,
+                min_p,
+                enable_thinking,
            )
        except Exception as e:
+            is_conn_error = isinstance(
+                e,
+                (
+                    httpx.ConnectError,
+                    httpx.RemoteProtocolError,
+                    ConnectionRefusedError,
+                    OSError,
+                ),
+            )
+            if (
+                is_conn_error
+                and attempt >= INSTANCE_HEALTH_CHECK_AFTER
+                and not await _check_instance_health(base_url)
+            ):
+                if instance_failed:
+                    instance_failed.set()
+                raise InstanceFailedError(
+                    f"Instance is down after {attempt + 1} failures: {e}"
+                ) from e
            if attempt < MAX_RETRIES - 1:
                wait = min(2**attempt, 60)
                logger.warning(
@@ -617,10 +694,16 @@ async def evaluate_benchmark(
    max_tokens: int,
    concurrency: int = 1,
    limit: int | None = None,
+    offset: int = 0,
    timeout: float | None = None,
    reasoning_effort: str | None = None,
    top_p: float | None = None,
+    top_k: int | None = None,
+    min_p: float | None = None,
+    enable_thinking: bool | None = None,
    difficulty: str | None = None,
+    checkpoint_path: Path | None = None,
+    release_version: str | None = None,
 ) -> list[QuestionResult]:
    """Run a benchmark. Returns per-question results."""
    import datasets
@@ -651,7 +734,21 @@ async def evaluate_benchmark(
        ds = ds.filter(lambda x: x["difficulty"] == difficulty)
        logger.info(f"Filtered to {len(ds)} {difficulty} problems")

+    if release_version and "release_version" in ds.column_names:
+        ds = ds.filter(lambda x: x["release_version"] == release_version)
+        logger.info(
+            f"Filtered to {len(ds)} problems with release_version={release_version}"
+        )
+
+    # Sort by question_id to match LCB runner ordering (scenario_router.py:60).
+    # This ensures [offset:offset+limit] slices select the same problems as vllm.
+    if "question_id" in ds.column_names:
+        ds = ds.sort("question_id")
+
    total = len(ds)
+    if offset > 0:
+        ds = ds.select(range(min(offset, total), total))
+        total = len(ds)
    if limit and limit < total:
        ds = ds.select(range(limit))
        total = limit
@@ -659,6 +756,13 @@ async def evaluate_benchmark(
    logger.info(
        f"Evaluating {benchmark_name}: {total} questions, concurrency={concurrency}, "
        f"temperature={temperature}, max_tokens={max_tokens}"
+        + (f", top_k={top_k}" if top_k is not None else "")
+        + (f", min_p={min_p}" if min_p is not None else "")
+        + (
+            f", enable_thinking={enable_thinking}"
+            if enable_thinking is not None
+            else ""
+        )
    )

    if config.kind == "code":
@@ -666,16 +770,64 @@ async def evaluate_benchmark(
            "Code benchmarks execute model-generated code. Use a sandboxed environment."
        )

+    # Load checkpoint for resume
+    checkpoint_data: dict[str | int, dict[str, Any]] = {}
+    if checkpoint_path and checkpoint_path.exists():
+        with open(checkpoint_path) as f:
+            for line in f:
+                entry = json.loads(line)
+                checkpoint_data[entry["question_id"]] = entry
+        logger.info(f"Loaded {len(checkpoint_data)} checkpointed results")
+
    semaphore = asyncio.Semaphore(concurrency)
+    instance_failed = asyncio.Event()
    results: list[QuestionResult | None] = [None] * total
    completed = 0
    lock = asyncio.Lock()

+    def _get_question_id(idx: int, doc: dict) -> str | int:
+        """Get a stable question ID for checkpointing."""
+        if benchmark_name == "livecodebench":
+            return doc.get("question_id", idx)
+        elif benchmark_name == "humaneval":
+            return doc.get("task_id", idx)
+        return idx
+
    async def process_question(
        idx: int, doc: dict, http_client: httpx.AsyncClient
    ) -> None:
        nonlocal completed
        system_msg = None
+        question_id = _get_question_id(idx, doc)
+
+        # Bail out early if instance is already dead
+        if instance_failed.is_set():
+            return
+
+        # Check checkpoint
+        if question_id in checkpoint_data:
+            cached = checkpoint_data[question_id]
+            results[idx] = QuestionResult(
+                question_id=question_id,
+                prompt=cached.get("prompt", ""),
+                response=cached.get("response", ""),
+                extracted_answer=cached.get("extracted_answer"),
+                gold_answer=cached.get("gold_answer", ""),
+                correct=cached.get("correct", False),
+                error=cached.get("error"),
+                prompt_tokens=cached.get("prompt_tokens", 0),
+                completion_tokens=cached.get("completion_tokens", 0),
+                reasoning_tokens=cached.get("reasoning_tokens", 0),
+                reasoning_content=cached.get("reasoning_content", ""),
+                finish_reason=cached.get("finish_reason", ""),
+                elapsed_s=cached.get("elapsed_s", 0.0),
+                power_watts=cached.get("power_watts", 0.0),
+                energy_joules=cached.get("energy_joules", 0.0),
+            )
+            async with lock:
+                completed += 1
+            logger.info(f"  [{completed}/{total}] {question_id} (cached)")
+            return

        if benchmark_name == "gpqa_diamond":
            prompt, gold = format_gpqa_question(doc, idx)
@@ -696,24 +848,50 @@ async def evaluate_benchmark(
            raise ValueError(f"Unknown benchmark: {benchmark_name}")

        async with semaphore:
+            if instance_failed.is_set():
+                return
            t0 = time.monotonic()
-            api_result = await call_with_retries(
-                http_client,
-                base_url,
-                model,
-                prompt,
-                temperature,
-                max_tokens,
-                timeout,
-                system_message=system_msg,
-                reasoning_effort=reasoning_effort,
-                top_p=top_p,
-            )
+            try:
+                # Race the API call against the instance_failed event
+                api_task = asyncio.create_task(
+                    call_with_retries(
+                        http_client,
+                        base_url,
+                        model,
+                        prompt,
+                        temperature,
+                        max_tokens,
+                        timeout,
+                        system_message=system_msg,
+                        reasoning_effort=reasoning_effort,
+                        top_p=top_p,
+                        top_k=top_k,
+                        min_p=min_p,
+                        enable_thinking=enable_thinking,
+                        instance_failed=instance_failed,
+                    )
+                )
+                failed_waiter = asyncio.create_task(instance_failed.wait())
+                done, pending = await asyncio.wait(
+                    [api_task, failed_waiter],
+                    return_when=asyncio.FIRST_COMPLETED,
+                )
+                for p in pending:
+                    p.cancel()
+                    with contextlib.suppress(asyncio.CancelledError):
+                        await p
+                if instance_failed.is_set() and api_task not in done:
+                    logger.error(f"Instance failed, aborting {question_id}")
+                    return
+                api_result = api_task.result()
+            except InstanceFailedError:
+                logger.error(f"Instance failed, skipping {question_id}")
+                return
            elapsed = time.monotonic() - t0

        if api_result is None:
            result = QuestionResult(
-                question_id=idx,
+                question_id=question_id,
                prompt=prompt,
                response="",
                extracted_answer=None,
@@ -728,13 +906,17 @@ async def evaluate_benchmark(
                "prompt_tokens": api_result.prompt_tokens,
                "completion_tokens": api_result.completion_tokens,
                "reasoning_tokens": api_result.reasoning_tokens,
+                "reasoning_content": api_result.reasoning_content,
+                "finish_reason": api_result.finish_reason,
                "elapsed_s": elapsed,
+                "power_watts": api_result.power_watts,
+                "energy_joules": api_result.energy_joules,
            }

            if config.kind == "mc":
                extracted = extract_mc_answer(response, valid_letters)
                result = QuestionResult(
-                    question_id=idx,
+                    question_id=question_id,
                    prompt=prompt,
                    response=response,
                    extracted_answer=extracted,
@@ -748,7 +930,7 @@ async def evaluate_benchmark(
                    check_aime_answer(extracted, int(gold)) if extracted else False
                )
                result = QuestionResult(
-                    question_id=idx,
+                    question_id=question_id,
                    prompt=prompt,
                    response=response,
                    extracted_answer=extracted,
@@ -762,7 +944,7 @@ async def evaluate_benchmark(
                code = extract_code_block(response, preserve_indent=keep_indent)
                if code is None:
                    result = QuestionResult(
-                        question_id=idx,
+                        question_id=question_id,
                        prompt=prompt,
                        response=response,
                        extracted_answer=None,
@@ -777,7 +959,7 @@ async def evaluate_benchmark(
                        code,
                    )
                    result = QuestionResult(
-                        question_id=idx,
+                        question_id=question_id,
                        prompt=prompt,
                        response=response,
                        extracted_answer="pass" if passed else "fail",
@@ -792,7 +974,7 @@ async def evaluate_benchmark(
                        exec_meta["sample"],
                    )
                    result = QuestionResult(
-                        question_id=idx,
+                        question_id=question_id,
                        prompt=prompt,
                        response=response,
                        extracted_answer="pass" if passed else "fail",
@@ -803,7 +985,7 @@ async def evaluate_benchmark(
                    )
                else:
                    result = QuestionResult(
-                        question_id=idx,
+                        question_id=question_id,
                        prompt=prompt,
                        response=response,
                        extracted_answer=None,
@@ -814,7 +996,7 @@ async def evaluate_benchmark(
                    )
            else:
                result = QuestionResult(
-                    question_id=idx,
+                    question_id=question_id,
                    prompt=prompt,
                    response=response,
                    extracted_answer=None,
@@ -826,24 +1008,82 @@ async def evaluate_benchmark(

        results[idx] = result

+        # Write checkpoint (skip infra failures so they get retried on resume,
+        # but keep wrong answers — they are legitimate results)
+        if checkpoint_path is not None and result.response:
+            _write_checkpoint(checkpoint_path, result)
+
        async with lock:
            completed += 1
            n = completed
-        if n % max(1, total // 20) == 0 or n == total:
-            correct_so_far = sum(1 for r in results if r is not None and r.correct)
-            answered = sum(1 for r in results if r is not None)
-            logger.info(
-                f"  [{n}/{total}] {correct_so_far}/{answered} correct "
-                f"({correct_so_far / max(answered, 1):.1%})"
-            )
+
+        # Log progress
+        thinking_info = ""
+        if result.reasoning_content:
+            thinking_info = f", {len(result.reasoning_content)} chars thinking"
+        logger.info(
+            f"  [{n}/{total}] {question_id}: {len(result.response)} chars{thinking_info}, "
+            f"tokens: {result.prompt_tokens}+{result.completion_tokens} "
+            f"[{result.finish_reason}]"
+            + (f" {result.extracted_answer}" if result.extracted_answer else "")
+        )
+
+    async def _health_monitor() -> None:
+        """Periodically check if the instance is still alive."""
+        # Wait a bit before first check to let things start
+        await asyncio.sleep(10)
+        while not instance_failed.is_set():
+            if not await _check_instance_health(base_url):
+                # Double-check to avoid false positives
+                await asyncio.sleep(2)
+                if not await _check_instance_health(base_url):
+                    logger.error("Health monitor: instance is down!")
+                    instance_failed.set()
+                    return
+            await asyncio.sleep(5)

    async with httpx.AsyncClient() as http_client:
+        monitor = asyncio.create_task(_health_monitor())
        tasks = [process_question(i, doc, http_client) for i, doc in enumerate(ds)]
        await asyncio.gather(*tasks)
+        monitor.cancel()
+        with contextlib.suppress(asyncio.CancelledError):
+            await monitor
+
+    if instance_failed.is_set():
+        completed_count = sum(1 for r in results if r is not None)
+        logger.error(
+            f"Instance failed! Completed {completed_count}/{total} problems. "
+            f"Checkpoint saved — restart to resume remaining problems."
+        )
+        raise InstanceFailedError("Instance failed during evaluation")

    return [r for r in results if r is not None]


+def _write_checkpoint(path: Path, result: QuestionResult) -> None:
+    """Append a single result to the JSONL checkpoint file."""
+    entry = {
+        "question_id": result.question_id,
+        "prompt": result.prompt,
+        "response": result.response,
+        "extracted_answer": result.extracted_answer,
+        "gold_answer": result.gold_answer,
+        "correct": result.correct,
+        "error": result.error,
+        "prompt_tokens": result.prompt_tokens,
+        "completion_tokens": result.completion_tokens,
+        "reasoning_tokens": result.reasoning_tokens,
+        "reasoning_content": result.reasoning_content,
+        "finish_reason": result.finish_reason,
+        "elapsed_s": round(result.elapsed_s, 2),
+        "power_watts": round(result.power_watts, 2),
+        "energy_joules": round(result.energy_joules, 2),
+    }
+    with open(path, "a") as f:
+        f.write(json.dumps(entry) + "\n")
+
+
 # ---------------------------------------------------------------------------
 # Results display
 # ---------------------------------------------------------------------------
@@ -866,6 +1106,8 @@ def print_results(
    total_elapsed = sum(r.elapsed_s for r in results)
    wall_clock = max(r.elapsed_s for r in results) if results else 0.0
    avg_gen_tps = total_completion_tokens / total_elapsed if total_elapsed > 0 else 0.0
+    total_energy = sum(r.energy_joules for r in results)
+    avg_power = sum(r.power_watts for r in results) / max(total, 1)

    label = f"[c={concurrency}] " if concurrency is not None else ""
    print(f"\n{label}{benchmark_name}: {correct}/{total} ({accuracy:.1%})")
@@ -877,6 +1119,10 @@ def print_results(
        f"  |  total time: {total_elapsed:.1f}s  wall clock: {wall_clock:.1f}s"
    )
    print(tok_line)
+    if total_energy > 0:
+        print(
+            f"  power: avg {avg_power:.1f}W  |  total energy: {total_energy:.1f}J ({total_energy / 3600:.2f}Wh)"
+        )
    if errors:
        print(f"  API errors: {errors}")
    if no_extract:
@@ -895,6 +1141,8 @@ def print_results(
        "total_elapsed_s": total_elapsed,
        "wall_clock_s": wall_clock,
        "avg_gen_tps": avg_gen_tps,
+        "avg_power_watts": avg_power,
+        "total_energy_joules": total_energy,
    }


@@ -1027,16 +1275,18 @@ def save_results(
    concurrency: int,
    results: list[QuestionResult],
    scores: dict[str, Any],
+    cluster: dict[str, Any] | None = None,
 ) -> Path:
    out_dir = Path(results_dir) / model.replace("/", "_") / benchmark_name
    out_dir.mkdir(parents=True, exist_ok=True)
    ts = time.strftime("%Y%m%d_%H%M%S")
    path = out_dir / f"c{concurrency}_{ts}.json"

-    data = {
+    data: dict[str, Any] = {
        "benchmark": benchmark_name,
        "model": model,
        "concurrency": concurrency,
+        **({"cluster": cluster} if cluster else {}),
        "scores": scores,
        "results": [
            {
@@ -1050,7 +1300,11 @@ def save_results(
                "prompt_tokens": r.prompt_tokens,
                "completion_tokens": r.completion_tokens,
                "reasoning_tokens": r.reasoning_tokens,
+                "reasoning_content": r.reasoning_content,
+                "finish_reason": r.finish_reason,
                "elapsed_s": round(r.elapsed_s, 2),
+                "power_watts": round(r.power_watts, 2),
+                "energy_joules": round(r.energy_joules, 2),
            }
            for r in results
        ],
@@ -1066,6 +1320,15 @@ def save_results(
 # ---------------------------------------------------------------------------


+def _checkpoint_path(
+    results_dir: str, benchmark: str, model: str, concurrency: int
+) -> Path:
+    """Return the JSONL checkpoint path for a benchmark run."""
+    out_dir = Path(results_dir) / model.replace("/", "_") / benchmark
+    out_dir.mkdir(parents=True, exist_ok=True)
+    return out_dir / f"c{concurrency}.checkpoint.jsonl"
+
+
 def parse_int_list(values: list[str]) -> list[int]:
    items: list[int] = []
    for v in values:
@@ -1093,6 +1356,12 @@ def main() -> int:
        default=None,
        help="Max questions per benchmark (for fast iteration).",
    )
+    ap.add_argument(
+        "--offset",
+        type=int,
+        default=0,
+        help="Skip first N questions (0-based).",
+    )

    reasoning_group = ap.add_mutually_exclusive_group()
    reasoning_group.add_argument(
@@ -1112,6 +1381,8 @@ def main() -> int:
        "--temperature", type=float, default=None, help="Override temperature."
    )
    ap.add_argument("--top-p", type=float, default=None, help="Override top_p.")
+    ap.add_argument("--top-k", type=int, default=None, help="Override top_k.")
+    ap.add_argument("--min-p", type=float, default=None, help="Override min_p.")
    ap.add_argument(
        "--max-tokens", type=int, default=None, help="Override max output tokens."
    )
@@ -1145,15 +1416,31 @@ def main() -> int:
        choices=["easy", "medium", "hard"],
        help="Filter by difficulty (livecodebench only). E.g. --difficulty hard",
    )
+    ap.add_argument(
+        "--release-version",
+        default=None,
+        help="LCB dataset release version (livecodebench only). E.g. release_v5",
+    )
    ap.add_argument(
        "--results-dir",
        default="eval_results",
        help="Directory for result JSON files (default: eval_results).",
    )
    ap.add_argument(
-        "--skip-instance-setup",
+        "--enable-thinking",
+        type=lambda v: v.lower() in ("true", "1", "yes"),
+        default=None,
+        help="Enable thinking mode for models that support it.",
+    )
+    ap.add_argument(
+        "--force",
        action="store_true",
-        help="Skip exo instance management (assumes model is already running).",
+        help="Discard any existing checkpoint and run from scratch.",
+    )
+    ap.add_argument(
+        "--keep-instance",
+        action="store_true",
+        help="Skip deleting the instance after eval (for chaining runs).",
    )

    args, _ = ap.parse_known_args()
@@ -1174,13 +1461,26 @@ def main() -> int:
    # Instance management
    client = ExoClient(args.host, args.port, timeout_s=args.timeout)
    instance_id: str | None = None
+    created_instance = False

-    if not args.skip_instance_setup:
-        short_id, full_model_id = resolve_model_short_id(
-            client,
-            args.model,
-            force_download=args.force_download,
-        )
+    _short_id, full_model_id = resolve_model_short_id(
+        client,
+        args.model,
+        force_download=args.force_download,
+    )
+
+    # Optionally reuse a running instance for this model
+    if args.reuse_instance:
+        existing = find_existing_instance(client, full_model_id)
+        if existing:
+            instance_id = existing
+            logger.info(f"Reusing existing instance {instance_id}")
+        else:
+            logger.warning(
+                "--reuse-instance: no existing instance found, creating a new one"
+            )
+
+    if instance_id is None:
        selected = settle_and_fetch_placements(
            client,
            full_model_id,
@@ -1195,7 +1495,7 @@ def main() -> int:
            key=lambda p: (
                str(p.get("instance_meta", "")),
                str(p.get("sharding", "")),
-                -nodes_used_in_instance(p["instance"]),
+                nodes_used_in_instance(p["instance"]),
            ),
            reverse=True,
        )
@@ -1222,6 +1522,18 @@ def main() -> int:
        if download_duration is not None:
            logger.info(f"Download: {download_duration:.1f}s")

+        # Delete any existing instances to free resources before placing
+        try:
+            state = client.request_json("GET", "/state")
+            for old_id in list(state.get("instances", {}).keys()):
+                logger.info(f"Deleting stale instance {old_id}")
+                with contextlib.suppress(ExoHttpError):
+                    client.request_json("DELETE", f"/instance/{old_id}")
+            if state.get("instances"):
+                time.sleep(2)
+        except Exception as e:
+            logger.warning(f"Failed to clean up stale instances: {e}")
+
        client.request_json("POST", "/instance", body={"instance": instance})
        try:
            wait_for_instance_ready(client, instance_id)
@@ -1231,8 +1543,9 @@ def main() -> int:
                client.request_json("DELETE", f"/instance/{instance_id}")
            return 1
        time.sleep(1)
-    else:
-        full_model_id = args.model
+        created_instance = True
+
+    cluster_snapshot = capture_cluster_snapshot(client)

    # Auto-detect reasoning from model config
    model_config = load_model_config(full_model_id)
@@ -1286,16 +1599,57 @@ def main() -> int:
        reasoning_effort = str(cfg["reasoning_effort"])
    else:
        reasoning_effort = "high" if is_reasoning else None
+
+    if args.top_k is not None:
+        top_k: int | None = args.top_k
+    elif "top_k" in cfg:
+        top_k = int(cfg["top_k"])
+    else:
+        top_k = None
+
+    if args.min_p is not None:
+        min_p: float | None = args.min_p
+    elif "min_p" in cfg:
+        min_p = float(cfg["min_p"])
+    else:
+        min_p = None
+
+    if args.enable_thinking is not None:
+        enable_thinking: bool | None = args.enable_thinking
+    elif "enable_thinking" in cfg:
+        enable_thinking = bool(cfg["enable_thinking"])
+    else:
+        enable_thinking = None
+
    base_url = f"http://{args.host}:{args.port}"

    logger.info(f"Model: {full_model_id}")
    logger.info(
        f"Settings: temperature={temperature}, max_tokens={max_tokens}, "
        + (f"top_p={top_p}, " if top_p is not None else "")
+        + (f"top_k={top_k}, " if top_k is not None else "")
+        + (f"min_p={min_p}, " if min_p is not None else "")
        + f"reasoning={'yes' if is_reasoning else 'no'}"
        + (f", reasoning_effort={reasoning_effort}" if reasoning_effort else "")
+        + (
+            f", enable_thinking={enable_thinking}"
+            if enable_thinking is not None
+            else ""
+        )
    )

+    # Common kwargs for evaluate_benchmark
+    eval_kwargs: dict[str, Any] = {
+        "reasoning_effort": reasoning_effort,
+        "top_p": top_p,
+        "top_k": top_k,
+        "min_p": min_p,
+        "enable_thinking": enable_thinking,
+        "difficulty": args.difficulty,
+        "offset": args.offset,
+        "release_version": args.release_version,
+    }
+
    try:
        if args.compare_concurrency:
            concurrency_levels = parse_int_list(args.compare_concurrency)
@@ -1304,6 +1658,11 @@ def main() -> int:
                for c in concurrency_levels:
                    logger.info(f"\n{'=' * 50}")
                    logger.info(f"Running {task_name} at concurrency={c}")
+                    checkpoint_path = _checkpoint_path(
+                        args.results_dir, task_name, full_model_id, c
+                    )
+                    if args.force and checkpoint_path.exists():
+                        checkpoint_path.unlink()
                    results = asyncio.run(
                        evaluate_benchmark(
                            task_name,
@@ -1314,9 +1673,8 @@ def main() -> int:
                            concurrency=c,
                            limit=args.limit,
                            timeout=args.request_timeout,
-                            reasoning_effort=reasoning_effort,
-                            top_p=top_p,
-                            difficulty=args.difficulty,
+                            checkpoint_path=checkpoint_path,
+                            **eval_kwargs,
                        )
                    )
                    if results:
@@ -1328,12 +1686,21 @@ def main() -> int:
                            c,
                            results,
                            scores,
+                            cluster=cluster_snapshot,
                        )
                        results_by_c[c] = results
+                    # Clean up checkpoint on success
+                    if checkpoint_path.exists():
+                        checkpoint_path.unlink()
                if len(results_by_c) >= 2:
                    print_comparison(task_name, results_by_c)
        else:
            for task_name in task_names:
+                checkpoint_path = _checkpoint_path(
+                    args.results_dir, task_name, full_model_id, args.num_concurrent
+                )
+                if args.force and checkpoint_path.exists():
+                    checkpoint_path.unlink()
                results = asyncio.run(
                    evaluate_benchmark(
                        task_name,
@@ -1344,9 +1711,8 @@ def main() -> int:
                        concurrency=args.num_concurrent,
                        limit=args.limit,
                        timeout=args.request_timeout,
-                        reasoning_effort=reasoning_effort,
-                        top_p=top_p,
-                        difficulty=args.difficulty,
+                        checkpoint_path=checkpoint_path,
+                        **eval_kwargs,
                    )
                )
                if results:
@@ -1358,15 +1724,27 @@ def main() -> int:
                        args.num_concurrent,
                        results,
                        scores,
+                        cluster=cluster_snapshot,
                    )
+                # Clean up checkpoint on success
+                if checkpoint_path.exists():
+                    checkpoint_path.unlink()
    finally:
-        if instance_id is not None:
-            try:
-                client.request_json("DELETE", f"/instance/{instance_id}")
-            except ExoHttpError as e:
-                if e.status != 404:
-                    raise
-            wait_for_instance_gone(client, instance_id)
+        if created_instance and instance_id is not None:
+            if args.keep_instance:
+                logger.info(f"Keeping instance {instance_id} (--keep-instance)")
+            else:
+                try:
+                    client.request_json("DELETE", f"/instance/{instance_id}")
+                except ExoHttpError as e:
+                    if e.status != 404:
+                        raise
+                try:
+                    wait_for_instance_gone(client, instance_id)
+                except TimeoutError:
+                    logger.warning(
+                        f"Timed out waiting for instance {instance_id} to be deleted"
+                    )

    return 0

@@ -6,6 +6,7 @@ import http.client
 import json
 import os
 import time
+from collections.abc import Iterator
 from typing import Any
 from urllib.parse import urlencode

@@ -69,6 +70,59 @@ class ExoClient:
    def post_bench_chat_completions(self, payload: dict[str, Any]) -> dict[str, Any]:
        return self.request_json("POST", "/bench/chat/completions", body=payload)

+    def stream_bench_chat_completions(self, payload: dict[str, Any]) -> Iterator[str]:
+        """POST /bench/chat/completions with stream=True, yielding raw SSE lines."""
+        payload = {**payload, "stream": True}
+        data = json.dumps(payload).encode("utf-8")
+        conn = http.client.HTTPConnection(self.host, self.port, timeout=self.timeout_s)
+        try:
+            conn.request(
+                "POST",
+                "/bench/chat/completions",
+                body=data,
+                headers={
+                    "Content-Type": "application/json",
+                    "Accept": "text/event-stream",
+                },
+            )
+            resp = conn.getresponse()
+            if resp.status >= 400:
+                raw = resp.read().decode("utf-8", errors="replace")
+                raise ExoHttpError(resp.status, resp.reason, raw[:300])
+            for line in resp:
+                yield line.decode("utf-8", errors="replace")
+        finally:
+            conn.close()
+
+    def get_state_path(self, path: str) -> Any:
+        try:
+            return self.request_json("GET", f"/state/{path}")
+        except ExoHttpError as e:
+            if e.status == 404:
+                return None
+            raise
+
+    def get_instance(self, instance_id: str) -> dict[str, Any] | None:
+        return self.get_state_path(f"instances/{instance_id}")
+
+    def get_runner(self, runner_id: str) -> dict[str, Any] | None:
+        return self.get_state_path(f"runners/{runner_id}")
+
+    def get_node_downloads(self, node_id: str) -> list[dict[str, Any]] | None:
+        return self.get_state_path(f"downloads/{node_id}")
+
+    def get_node_disk(self, node_id: str) -> dict[str, Any] | None:
+        return self.get_state_path(f"nodeDisk/{node_id}")
+
+    def get_node_system(self, node_id: str) -> dict[str, Any] | None:
+        return self.get_state_path(f"nodeSystem/{node_id}")
+
+    def get_node_identities(self) -> dict[str, Any] | None:
+        return self.get_state_path("nodeIdentities")
+
+    def get_topology(self) -> dict[str, Any] | None:
+        return self.get_state_path("topology")
+

 def unwrap_instance(instance: dict[str, Any]) -> dict[str, Any]:
    if len(instance) != 1:
@@ -97,6 +151,11 @@ def runner_ids_from_instance(instance: dict[str, Any]) -> list[str]:
    return list(runner_to_shard.keys())


+def node_ids_from_instance(instance: dict[str, Any]) -> list[str]:
+    inner = unwrap_instance(instance)
+    return list(inner["shardAssignments"]["nodeToRunner"].keys())
+
+
 def runner_ready(runner: dict[str, Any]) -> bool:
    return "RunnerReady" in runner

@@ -116,13 +175,12 @@ def wait_for_instance_ready(
 ) -> None:
    start_time = time.time()
    instance_existed = False
+    last_loaded: dict[str, int] = {}
    while time.time() - start_time < timeout:
-        state = client.request_json("GET", "/state")
-        instances = state.get("instances", {})
+        instance = client.get_instance(instance_id)

-        if instance_id not in instances:
+        if instance is None:
            if instance_existed:
-                # Instance was deleted after being created - likely due to runner failure
                raise RuntimeError(
                    f"Instance {instance_id} was deleted (runner may have failed)"
                )
@@ -130,18 +188,25 @@ def wait_for_instance_ready(
            continue

        instance_existed = True
-        instance = instances[instance_id]
-        runner_ids = runner_ids_from_instance(instance)
-        runners = state.get("runners", {})
+        rids = runner_ids_from_instance(instance)

-        # Check for failed runners first
-        for rid in runner_ids:
-            runner = runners.get(rid, {})
+        all_ready = True
+        for rid in rids:
+            runner = client.get_runner(rid) or {}
            if runner_failed(runner):
                error_msg = get_runner_failed_message(runner) or "Unknown error"
                raise RuntimeError(f"Runner {rid} failed: {error_msg}")
+            if "RunnerLoading" in runner:
+                loading = runner["RunnerLoading"]
+                loaded = loading.get("layersLoaded", 0)
+                total = loading.get("totalLayers", 0)
+                if total > 0 and last_loaded.get(rid) != loaded:
+                    last_loaded[rid] = loaded
+                    logger.debug(f"Runner {rid}: loading layers {loaded}/{total}")
+            if not runner_ready(runner):
+                all_ready = False

-        if all(runner_ready(runners.get(rid, {})) for rid in runner_ids):
+        if all_ready:
            return

        time.sleep(0.1)
@@ -165,6 +230,23 @@ def wait_for_instance_gone(
    raise TimeoutError(f"Instance {instance_id} did not get deleted within {timeout=}")


+def capture_cluster_snapshot(client: ExoClient) -> dict[str, Any]:
+    snapshot: dict[str, Any] = {}
+    identities = client.get_node_identities()
+    if identities:
+        snapshot["nodeIdentities"] = identities
+    topology = client.get_topology()
+    if topology:
+        snapshot["topology"] = topology
+    node_memory = client.get_state_path("nodeMemory")
+    if node_memory:
+        snapshot["nodeMemory"] = node_memory
+    node_system = client.get_state_path("nodeSystem")
+    if node_system:
+        snapshot["nodeSystem"] = node_system
+    return snapshot
+
+
 def resolve_model_short_id(
    client: ExoClient, model_arg: str, *, force_download: bool = False
 ) -> tuple[str, str]:
@@ -211,11 +293,15 @@ def sharding_filter(sharding: str, wanted: str) -> bool:


 def fetch_and_filter_placements(
-    client: ExoClient, full_model_id: str, args: argparse.Namespace
+    client: ExoClient,
+    full_model_id: str,
+    args: argparse.Namespace,
+    node_id: str | None = None,
 ) -> list[dict[str, Any]]:
-    previews_resp = client.request_json(
-        "GET", "/instance/previews", params={"model_id": full_model_id}
-    )
+    params: dict[str, str] = {"model_id": full_model_id}
+    if node_id is not None:
+        params["node_ids"] = node_id
+    previews_resp = client.request_json("GET", "/instance/previews", params=params)
    previews = previews_resp.get("previews") or []

    selected: list[dict[str, Any]] = []
@@ -275,8 +361,9 @@ def settle_and_fetch_placements(
    full_model_id: str,
    args: argparse.Namespace,
    settle_timeout: float = 0,
+    node_id: str | None = None,
 ) -> list[dict[str, Any]]:
-    selected = fetch_and_filter_placements(client, full_model_id, args)
+    selected = fetch_and_filter_placements(client, full_model_id, args, node_id=node_id)

    if not selected and settle_timeout > 0:
        backoff = _SETTLE_INITIAL_BACKOFF_S
@@ -289,7 +376,9 @@ def settle_and_fetch_placements(
            )
            time.sleep(min(backoff, remaining))
            backoff = min(backoff * _SETTLE_BACKOFF_MULTIPLIER, _SETTLE_MAX_BACKOFF_S)
-            selected = fetch_and_filter_placements(client, full_model_id, args)
+            selected = fetch_and_filter_placements(
+                client, full_model_id, args, node_id=node_id
+            )

    return selected

@@ -326,16 +415,11 @@ def run_planning_phase(
    node_ids = list(inner["shardAssignments"]["nodeToRunner"].keys())
    runner_to_shard = inner["shardAssignments"]["runnerToShard"]

-    state = client.request_json("GET", "/state")
-    downloads = state.get("downloads", {})
-    node_disk = state.get("nodeDisk", {})
-
    needs_download = False

    for node_id in node_ids:
-        node_downloads = downloads.get(node_id, [])
+        node_downloads = client.get_node_downloads(node_id) or []

-        # Check if model already downloaded on this node
        already_downloaded = any(
            "DownloadCompleted" in p
            and unwrap_instance(p["DownloadCompleted"]["shardMetadata"])["modelCard"][
@@ -349,8 +433,7 @@ def run_planning_phase(

        needs_download = True

-        # Wait for disk info if settle_deadline is set
-        disk_info = node_disk.get(node_id, {})
+        disk_info = client.get_node_disk(node_id) or {}
        backoff = _SETTLE_INITIAL_BACKOFF_S
        while not disk_info and settle_deadline and time.monotonic() < settle_deadline:
            remaining = settle_deadline - time.monotonic()
@@ -359,9 +442,7 @@ def run_planning_phase(
            )
            time.sleep(min(backoff, remaining))
            backoff = min(backoff * _SETTLE_BACKOFF_MULTIPLIER, _SETTLE_MAX_BACKOFF_S)
-            state = client.request_json("GET", "/state")
-            node_disk = state.get("nodeDisk", {})
-            disk_info = node_disk.get(node_id, {})
+            disk_info = client.get_node_disk(node_id) or {}

        if not disk_info:
            logger.warning(f"No disk info for {node_id}, skipping space check")
@@ -377,7 +458,6 @@ def run_planning_phase(
                f"have {avail // (1024**3)}GB. Use --danger-delete-downloads to free space."
            )

-        # Delete from smallest to largest (skip read-only models from EXO_MODELS_PATH)
        completed = [
            (
                unwrap_instance(p["DownloadCompleted"]["shardMetadata"])["modelCard"][
@@ -414,24 +494,22 @@ def run_planning_phase(
        )
        logger.info(f"Started download on {node_id}")

-    # Wait for downloads
-    start = time.time()
-    while time.time() - start < timeout:
-        state = client.request_json("GET", "/state")
-        downloads = state.get("downloads", {})
+    # Wait for downloads (no timeout — poll until complete or failed)
+    while True:
        all_done = True
        for node_id in node_ids:
+            node_downloads = client.get_node_downloads(node_id) or []
            done = any(
                "DownloadCompleted" in p
                and unwrap_instance(p["DownloadCompleted"]["shardMetadata"])[
                    "modelCard"
                ]["modelId"]
                == full_model_id
-                for p in downloads.get(node_id, [])
+                for p in node_downloads
            )
            failed = [
                p["DownloadFailed"]["errorMessage"]
-                for p in downloads.get(node_id, [])
+                for p in node_downloads
                if "DownloadFailed" in p
                and unwrap_instance(p["DownloadFailed"]["shardMetadata"])["modelCard"][
                    "modelId"
@@ -442,13 +520,49 @@ def run_planning_phase(
                raise RuntimeError(f"Download failed on {node_id}: {failed[0]}")
            if not done:
                all_done = False
+                ongoing = [
+                    p
+                    for p in node_downloads
+                    if "DownloadOngoing" in p
+                    and unwrap_instance(p["DownloadOngoing"]["shardMetadata"])[
+                        "modelCard"
+                    ]["modelId"]
+                    == full_model_id
+                ]
+                if ongoing:
+                    prog = ongoing[0]["DownloadOngoing"]["downloadProgress"]
+                    speed_mb = prog.get("speed", 0) / (1024 * 1024)
+                    eta_s = prog.get("etaMs", 0) / 1000
+                    dl_bytes = prog.get("downloaded", {}).get("inBytes", 0)
+                    total_bytes = prog.get("total", {}).get("inBytes", 0)
+                    pct = (dl_bytes / total_bytes * 100) if total_bytes else 0
+                    logger.info(
+                        f"Downloading on {node_id}: {pct:.1f}% @ {speed_mb:.1f} MB/s, "
+                        f"ETA {eta_s:.0f}s "
+                        f"({prog.get('completedFiles', 0)}/{prog.get('totalFiles', 0)} files)"
+                    )
        if all_done:
            if download_t0 is not None:
                return time.perf_counter() - download_t0
            return None
-        time.sleep(1)
+        time.sleep(10)

-    raise TimeoutError("Downloads did not complete in time")
+
+def find_existing_instance(client: ExoClient, model_id: str) -> str | None:
+    """Find an existing running instance for the given model."""
+    try:
+        state = client.request_json("GET", "/state")
+    except Exception:
+        return None
+    for inst_id, inst in state.get("instances", {}).items():
+        # Instance structure is nested: {"MlxJacclInstance": {"shardAssignments": {"modelId": ...}}}
+        for _inst_type, inner in inst.items():
+            if not isinstance(inner, dict):
+                continue
+            sa = inner.get("shardAssignments", {})
+            if sa.get("modelId") == model_id:
+                return inst_id
+    return None


 def add_common_instance_args(ap: argparse.ArgumentParser) -> None:
@@ -496,7 +610,7 @@ def add_common_instance_args(ap: argparse.ArgumentParser) -> None:
    ap.add_argument(
        "--settle-timeout",
        type=float,
-        default=0,
+        default=60.0,
        help="Max seconds to wait for the cluster to produce valid placements (0 = try once).",
    )
    ap.add_argument(
@@ -504,3 +618,8 @@ def add_common_instance_args(ap: argparse.ArgumentParser) -> None:
        action="store_true",
        help="Delete existing models from smallest to largest to make room for benchmark model.",
    )
+    ap.add_argument(
+        "--reuse-instance",
+        action="store_true",
+        help="Reuse an existing running instance for this model instead of creating a new one.",
+    )
@@ -0,0 +1,36 @@
+# Prefill/Decode disaggregation benchmark config.
+#
+# Top-level keys are bench-wide. [prefill] and [decode] sections set per-side
+# placement filters and (optionally) per-side model.
+#
+# Example:
+#   uv run python bench/prefill_decode_bench.py --config bench/prefill-decode.toml
+
+host = "james"
+port = 52415
+timeout = 7200.0
+settle_timeout = 60.0
+
+# Workload
+pp = [4096]
+tg = [512]
+repeat = 1
+warmup = 0
+
+json_out = "bench/prefill_decode_results.json"
+
+[prefill]
+model = "mlx-community/gpt-oss-20b-MXFP4-Q8"
+node = "mike"
+instance_meta = "ring"
+sharding = "pipeline"
+min_nodes = 1
+max_nodes = 1
+
+[decode]
+model = "mlx-community/gpt-oss-20b-MXFP4-Q8"
+node = "james"
+instance_meta = "ring"
+sharding = "pipeline"
+min_nodes = 1
+max_nodes = 1
@@ -0,0 +1,785 @@
+# type: ignore
+#!/usr/bin/env python3
+"""Disaggregated prefill-decode benchmark for exo (MLX → MLX).
+
+Spins up two MLX instances on the cluster, marks one as Prefill source and
+the other as Decode target via /v1/instance-links, then sends chat
+completions to the API. The master routes the request to the decode
+instance and stamps `prefill_endpoint` pointing at the prefill instance —
+the worker decides per-request whether to ship prefill remotely
+(uncached_count > REMOTE_PREFILL_MIN_TOKENS).
+
+Usage:
+    uv run python bench/prefill_decode_bench.py --model <id> --pp 2048,8192 --tg 128
+    uv run python bench/prefill_decode_bench.py --model <id> --pp 4096 --tg 128 --repeat 3
+    uv run python bench/prefill_decode_bench.py --model <id> --pp 2048 --tg 128 --dry-run
+"""
+
+from __future__ import annotations
+
+import argparse
+import contextlib
+import copy
+import itertools
+import json
+import sys
+import time
+import tomllib
+from pathlib import Path
+from statistics import mean
+from typing import Any
+
+from exo_bench import (
+    PromptSizer,
+    format_peak_memory,
+    load_tokenizer_for_bench,
+    parse_int_list,
+)
+from harness import (
+    ExoClient,
+    ExoHttpError,
+    add_common_instance_args,
+    instance_id_from_instance,
+    node_ids_from_instance,
+    nodes_used_in_instance,
+    resolve_model_short_id,
+    run_planning_phase,
+    settle_and_fetch_placements,
+    unwrap_instance,
+    wait_for_instance_gone,
+    wait_for_instance_ready,
+)
+from loguru import logger
+
+
+def _node_id_to_friendly(client: ExoClient) -> dict[str, str]:
+    identities = client.get_node_identities() or {}
+    out: dict[str, str] = {}
+    for node_id, identity in identities.items():
+        if isinstance(identity, dict):
+            name = identity.get("friendlyName") or identity.get("friendly_name")
+            if isinstance(name, str):
+                out[str(node_id)] = name
+    return out
+
+
+def _placement_node_friendly_names(
+    placement: dict[str, Any], id_to_friendly: dict[str, str]
+) -> list[str]:
+    instance = placement["instance"]
+    return [id_to_friendly.get(nid, nid) for nid in node_ids_from_instance(instance)]
+
+
+def _filter_by_node(
+    placements: list[dict[str, Any]],
+    friendly_name: str,
+    id_to_friendly: dict[str, str],
+) -> list[dict[str, Any]]:
+    target = friendly_name.lower()
+    matched: list[dict[str, Any]] = []
+    for p in placements:
+        names = [n.lower() for n in _placement_node_friendly_names(p, id_to_friendly)]
+        if any(target == n or target in n for n in names):
+            matched.append(p)
+    return matched
+
+
+def _node_id_by_friendly(id_to_friendly: dict[str, str], target: str) -> str | None:
+    target_lc = target.lower()
+    for nid, name in id_to_friendly.items():
+        if target_lc == name.lower() or target_lc in name.lower():
+            return nid
+    return None
+
+
+def _load_toml(path: str) -> dict[str, Any]:
+    with Path(path).open("rb") as f:
+        return tomllib.load(f)
+
+
+_TOP_LEVEL_TOML_KEYS = {
+    "host",
+    "port",
+    "timeout",
+    "settle_timeout",
+    "model",
+    "pp",
+    "tg",
+    "repeat",
+    "warmup",
+    "json_out",
+    "instance_meta",
+    "sharding",
+    "min_nodes",
+    "max_nodes",
+    "force_download",
+    "danger_delete_downloads",
+    "all_combinations",
+}
+
+
+def _inject_toml_into_argv() -> None:
+    """If --config X is in sys.argv, pre-load it and inject required CLI args
+    (--model, --pp, --tg) so argparse's required=True checks pass."""
+    argv = sys.argv
+    if "--config" not in argv:
+        return
+    idx = argv.index("--config")
+    if idx + 1 >= len(argv):
+        return
+    cfg_path = argv[idx + 1]
+    cfg = _load_toml(cfg_path)
+    decode = cfg.get("decode", {})
+
+    def _has(flag: str) -> bool:
+        return any(a == flag or a.startswith(flag + "=") for a in argv)
+
+    # --model: prefer top-level, then [decode].model
+    if not _has("--model"):
+        model = cfg.get("model") or decode.get("model")
+        if model:
+            argv += ["--model", str(model)]
+    if not _has("--pp"):
+        pp = cfg.get("pp")
+        if pp:
+            argv += (
+                ["--pp", *(str(x) for x in pp)]
+                if isinstance(pp, list)
+                else [
+                    "--pp",
+                    str(pp),
+                ]
+            )
+    if not _has("--tg"):
+        tg = cfg.get("tg")
+        if tg:
+            argv += (
+                ["--tg", *(str(x) for x in tg)]
+                if isinstance(tg, list)
+                else [
+                    "--tg",
+                    str(tg),
+                ]
+            )
+
+
+def _merge_toml_into_args(args: argparse.Namespace, cfg: dict[str, Any]) -> None:
+    """Apply top-level toml keys onto args namespace where args has a default."""
+    for key, value in cfg.items():
+        if key in {"prefill", "decode"}:
+            continue
+        if key not in _TOP_LEVEL_TOML_KEYS:
+            continue
+        attr = key
+        current = getattr(args, attr, None)
+        if current in (None, [], False):
+            setattr(args, attr, value)
+
+
+def _side_args(
+    base: argparse.Namespace, overrides: dict[str, Any]
+) -> argparse.Namespace:
+    out = copy.copy(base)
+    for k in (
+        "instance_meta",
+        "sharding",
+        "min_nodes",
+        "max_nodes",
+        "skip_pipeline_jaccl",
+        "skip_tensor_ring",
+    ):
+        if k in overrides:
+            setattr(out, k, overrides[k])
+    return out
+
+
+def _pick_two_distinct_placements(
+    placements: list[dict[str, Any]],
+) -> tuple[dict[str, Any], dict[str, Any]] | None:
+    if len(placements) < 2:
+        return None
+    seen_nodes: set[tuple[str, ...]] = set()
+    chosen: list[dict[str, Any]] = []
+    for p in placements:
+        nodes = tuple(sorted(str(n) for n in p.get("nodes", [])))
+        if nodes in seen_nodes:
+            continue
+        seen_nodes.add(nodes)
+        chosen.append(p)
+        if len(chosen) == 2:
+            return chosen[0], chosen[1]
+    return None
+
+
+def _create_instance_link(
+    client: ExoClient,
+    prefill_instance_id: str,
+    decode_instance_id: str,
+) -> str:
+    out = client.request_json(
+        "POST",
+        "/v1/instance-links",
+        body={
+            "prefill_instances": [prefill_instance_id],
+            "decode_instances": [decode_instance_id],
+        },
+    )
+    return str(out.get("commandId", ""))
+
+
+def _list_instance_links(client: ExoClient) -> list[dict[str, Any]]:
+    out = client.request_json("GET", "/v1/instance-links")
+    return out if isinstance(out, list) else []
+
+
+def _delete_instance_link(client: ExoClient, link_id: str) -> None:
+    client.request_json("DELETE", f"/v1/instance-links/{link_id}")
+
+
+def run_one(
+    client: ExoClient,
+    model_id: str,
+    pp_hint: int,
+    tg: int,
+    prompt_sizer: PromptSizer,
+) -> tuple[dict[str, Any], int]:
+    content, pp_tokens = prompt_sizer.build(pp_hint)
+    payload: dict[str, Any] = {
+        "model": model_id,
+        "messages": [{"role": "user", "content": content}],
+        "stream": False,
+        "max_tokens": tg,
+    }
+
+    t0 = time.perf_counter()
+    out = client.post_bench_chat_completions(payload)
+    elapsed = time.perf_counter() - t0
+
+    stats = out.get("generation_stats")
+    choices = out.get("choices") or [{}]
+    message = choices[0].get("message", {}) if choices else {}
+    text = message.get("content") or ""
+    preview = text[:200] if text else ""
+
+    return {
+        "elapsed_s": elapsed,
+        "output_text_preview": preview,
+        "stats": stats,
+    }, pp_tokens
+
+
+def _run_phase(
+    *,
+    client: ExoClient,
+    label: str,
+    pp_tg_pairs: list[tuple[int, int]],
+    model_id: str,
+    prompt_sizer: PromptSizer,
+    warmup: int,
+    repeat: int,
+    common_meta: dict[str, Any],
+) -> list[dict[str, Any]]:
+    logger.info(f"=== phase: {label} (model={model_id}) ===")
+    rows: list[dict[str, Any]] = []
+    for i in range(warmup):
+        run_one(client, model_id, pp_tg_pairs[0][0], pp_tg_pairs[0][1], prompt_sizer)
+        logger.debug(f"  warmup {i + 1}/{warmup} done")
+
+    for pp, tg in pp_tg_pairs:
+        logger.info(f"--- {label}: pp={pp} tg={tg} ---")
+        runs: list[dict[str, Any]] = []
+        for r in range(repeat):
+            time.sleep(2)
+            try:
+                row, actual_pp_tokens = run_one(client, model_id, pp, tg, prompt_sizer)
+            except Exception as e:
+                logger.error(e)
+                continue
+            row.update(common_meta)
+            row.update(
+                {
+                    "phase": label,
+                    "phase_model_id": model_id,
+                    "pp_tokens": actual_pp_tokens,
+                    "tg": tg,
+                    "repeat_index": r,
+                }
+            )
+            runs.append(row)
+            rows.append(row)
+
+        if runs:
+            prompt_tps = mean(x["stats"]["prompt_tps"] for x in runs)
+            gen_tps = mean(x["stats"]["generation_tps"] for x in runs)
+            ptok = mean(x["stats"]["prompt_tokens"] for x in runs)
+            gtok = mean(x["stats"]["generation_tokens"] for x in runs)
+            peak = mean(x["stats"]["peak_memory_usage"]["inBytes"] for x in runs)
+            avg_elapsed = mean(x["elapsed_s"] for x in runs)
+            logger.info(
+                f"[{label}] prompt_tps={prompt_tps:.2f} gen_tps={gen_tps:.2f}    "
+                f"prompt_tokens={ptok} gen_tokens={gtok}    "
+                f"peak_memory={format_peak_memory(peak)}    "
+                f"avg_elapsed={avg_elapsed:.2f}s"
+            )
+        time.sleep(2)
+    return rows
+
+
+def _summarise(rows: list[dict[str, Any]]) -> dict[tuple[int, int], dict[str, float]]:
+    grouped: dict[tuple[int, int], list[dict[str, Any]]] = {}
+    for r in rows:
+        key = (int(r["pp_tokens"]), int(r["tg"]))
+        grouped.setdefault(key, []).append(r)
+    out: dict[tuple[int, int], dict[str, float]] = {}
+    for key, runs in grouped.items():
+        out[key] = {
+            "prompt_tps": mean(x["stats"]["prompt_tps"] for x in runs),
+            "gen_tps": mean(x["stats"]["generation_tps"] for x in runs),
+            "elapsed_s": mean(x["elapsed_s"] for x in runs),
+        }
+    return out
+
+
+def _print_diff(
+    disagg_rows: list[dict[str, Any]],
+    decode_alone_rows: list[dict[str, Any]],
+    prefill_alone_rows: list[dict[str, Any]],
+) -> None:
+    disagg = _summarise(disagg_rows)
+    decode_alone = _summarise(decode_alone_rows)
+    prefill_alone = _summarise(prefill_alone_rows)
+    keys = set(disagg.keys()) | set(decode_alone.keys()) | set(prefill_alone.keys())
+
+    width = 64
+    for key in sorted(keys):
+        pp, tg = key
+        logger.info("─" * width)
+        logger.info(f"  pp={pp}  tg={tg}")
+        logger.info("─" * width)
+        logger.info(
+            f"  {'phase':<16} {'elapsed':>10}  {'prompt_tps':>11}  {'gen_tps':>9}"
+        )
+        for label, summary in (
+            ("disaggregated", disagg.get(key)),
+            ("decode_alone", decode_alone.get(key)),
+            ("prefill_alone", prefill_alone.get(key)),
+        ):
+            if summary is None:
+                logger.info(f"  {label:<16} {'—':>10}  {'—':>11}  {'—':>9}")
+                continue
+            logger.info(
+                f"  {label:<16} "
+                f"{summary['elapsed_s']:>9.2f}s  "
+                f"{summary['prompt_tps']:>11.1f}  "
+                f"{summary['gen_tps']:>9.2f}"
+            )
+
+        d = disagg.get(key)
+        da = decode_alone.get(key)
+        pa = prefill_alone.get(key)
+        if d and da and d["elapsed_s"] > 0:
+            logger.info(
+                f"  speedup vs decode_alone:  {da['elapsed_s'] / d['elapsed_s']:.2f}x"
+            )
+        if d and pa and d["elapsed_s"] > 0:
+            logger.info(
+                f"  speedup vs prefill_alone: {pa['elapsed_s'] / d['elapsed_s']:.2f}x"
+            )
+    logger.info("─" * width)
+
+
+def main() -> int:
+    _inject_toml_into_argv()
+    ap = argparse.ArgumentParser(
+        prog="prefill-decode-bench",
+        description="Benchmark MLX-MLX disaggregated prefill/decode via instance links.",
+    )
+    add_common_instance_args(ap)
+    ap.add_argument(
+        "--pp",
+        nargs="+",
+        required=True,
+        help="Prompt-size hints (ints, must be >1000). Accepts commas.",
+    )
+    ap.add_argument(
+        "--tg",
+        nargs="+",
+        required=True,
+        help="Generation lengths (ints). Accepts commas.",
+    )
+    ap.add_argument(
+        "--repeat", type=int, default=1, help="Repetitions per (pp,tg) pair."
+    )
+    ap.add_argument(
+        "--warmup",
+        type=int,
+        default=0,
+        help="Warmup runs (uses first pp/tg).",
+    )
+    ap.add_argument(
+        "--json-out",
+        default="bench/prefill_decode_results.json",
+        help="Write raw per-run results JSON to this path.",
+    )
+    ap.add_argument("--stdout", action="store_true", help="Write results to stdout")
+    ap.add_argument(
+        "--dry-run", action="store_true", help="List selected placements and exit."
+    )
+    ap.add_argument(
+        "--all-combinations",
+        action="store_true",
+        help="Force all pp×tg combinations even when lists have equal length.",
+    )
+    ap.add_argument(
+        "--prefill-model",
+        default=None,
+        help="Model id for the prefill instance. Defaults to --model.",
+    )
+    ap.add_argument(
+        "--prefill-node",
+        default=None,
+        help="friendly_name of the node hosting the prefill instance.",
+    )
+    ap.add_argument(
+        "--decode-node",
+        default=None,
+        help="friendly_name of the node hosting the decode instance.",
+    )
+    ap.add_argument(
+        "--config",
+        default=None,
+        help="TOML config file. CLI flags override toml values.",
+    )
+    ap.add_argument(
+        "--compare-baseline",
+        action="store_true",
+        help="Also run each (pp,tg) pair without the prefill/decode link "
+        "(decode instance does its own prefill) and report the diff.",
+    )
+    args = ap.parse_args()
+    cfg = _load_toml(args.config) if args.config else {}
+    _merge_toml_into_args(args, cfg)
+    prefill_overrides = cfg.get("prefill", {}) if cfg else {}
+    decode_overrides = cfg.get("decode", {}) if cfg else {}
+    if args.prefill_model is None and "model" in prefill_overrides:
+        args.prefill_model = prefill_overrides["model"]
+    if args.prefill_node is None and "node" in prefill_overrides:
+        args.prefill_node = prefill_overrides["node"]
+    if args.decode_node is None and "node" in decode_overrides:
+        args.decode_node = decode_overrides["node"]
+    if "model" in decode_overrides and not args.model:
+        args.model = decode_overrides["model"]
+
+    pp_list = parse_int_list(args.pp)
+    tg_list = parse_int_list(args.tg)
+    if not pp_list or not tg_list:
+        logger.error("pp and tg lists must be non-empty")
+        return 2
+    for pp in pp_list:
+        if pp <= 1000:
+            logger.error(
+                f"pp={pp} must be >1000 (remote prefill triggers when uncached >1000)"
+            )
+            return 2
+    if args.repeat <= 0:
+        logger.error("--repeat must be >= 1")
+        return 2
+
+    use_combinations = args.all_combinations or len(pp_list) != len(tg_list)
+    if use_combinations:
+        logger.info(
+            f"pp/tg mode: combinations (product) — {len(pp_list) * len(tg_list)} pairs"
+        )
+    else:
+        logger.info(f"pp/tg mode: tandem (zip) — {len(pp_list)} pairs")
+
+    client = ExoClient(args.host, args.port, timeout_s=args.timeout)
+
+    decode_short_id, decode_full_id = resolve_model_short_id(
+        client, args.model, force_download=args.force_download
+    )
+    if args.prefill_model:
+        prefill_short_id, prefill_full_id = resolve_model_short_id(
+            client, args.prefill_model, force_download=args.force_download
+        )
+    else:
+        prefill_short_id, prefill_full_id = decode_short_id, decode_full_id
+
+    tokenizer = load_tokenizer_for_bench(decode_full_id)
+    if tokenizer is None:
+        raise RuntimeError("[prefill-decode-bench] decode tokenizer load failed")
+    try:
+        decode_prompt_sizer = PromptSizer(tokenizer)
+    except Exception:
+        logger.error("[prefill-decode-bench] decode prompt sizing failed")
+        raise
+
+    if prefill_full_id == decode_full_id:
+        prefill_prompt_sizer = decode_prompt_sizer
+    else:
+        prefill_tokenizer = load_tokenizer_for_bench(prefill_full_id)
+        if prefill_tokenizer is None:
+            raise RuntimeError("[prefill-decode-bench] prefill tokenizer load failed")
+        prefill_prompt_sizer = PromptSizer(prefill_tokenizer)
+
+    id_to_friendly = _node_id_to_friendly(client)
+
+    prefill_args = _side_args(args, prefill_overrides)
+    decode_args = _side_args(args, decode_overrides)
+
+    if prefill_full_id == decode_full_id and prefill_overrides == decode_overrides:
+        placements = settle_and_fetch_placements(
+            client, decode_full_id, args, settle_timeout=args.settle_timeout
+        )
+        prefill_candidates = (
+            _filter_by_node(placements, args.prefill_node, id_to_friendly)
+            if args.prefill_node
+            else placements
+        )
+        decode_candidates = (
+            _filter_by_node(placements, args.decode_node, id_to_friendly)
+            if args.decode_node
+            else placements
+        )
+        if args.prefill_node and not prefill_candidates:
+            logger.error(f"No placement on prefill node {args.prefill_node!r}.")
+            return 1
+        if args.decode_node and not decode_candidates:
+            logger.error(f"No placement on decode node {args.decode_node!r}.")
+            return 1
+        if args.prefill_node and args.decode_node:
+            prefill_p = prefill_candidates[0]
+            decode_p = decode_candidates[0]
+        else:
+            pair = _pick_two_distinct_placements(placements)
+            if pair is None:
+                logger.error(
+                    "Need at least two distinct-node MLX placements for the same model."
+                )
+                return 1
+            prefill_p, decode_p = pair
+            if args.prefill_node:
+                prefill_p = prefill_candidates[0]
+            if args.decode_node:
+                decode_p = decode_candidates[0]
+    else:
+        prefill_node_id = (
+            _node_id_by_friendly(id_to_friendly, args.prefill_node)
+            if args.prefill_node
+            else None
+        )
+        decode_node_id = (
+            _node_id_by_friendly(id_to_friendly, args.decode_node)
+            if args.decode_node
+            else None
+        )
+        if args.prefill_node and prefill_node_id is None:
+            logger.error(f"Unknown node {args.prefill_node!r}.")
+            return 1
+        if args.decode_node and decode_node_id is None:
+            logger.error(f"Unknown node {args.decode_node!r}.")
+            return 1
+        prefill_placements = settle_and_fetch_placements(
+            client,
+            prefill_full_id,
+            prefill_args,
+            settle_timeout=args.settle_timeout,
+            node_id=prefill_node_id,
+        )
+        decode_placements = settle_and_fetch_placements(
+            client,
+            decode_full_id,
+            decode_args,
+            settle_timeout=args.settle_timeout,
+            node_id=decode_node_id,
+        )
+        if not prefill_placements:
+            logger.error(
+                f"No placement found for prefill model {prefill_full_id}"
+                f"{f' on node {args.prefill_node!r}' if args.prefill_node else ''}."
+            )
+            return 1
+        if not decode_placements:
+            logger.error(
+                f"No placement found for decode model {decode_full_id}"
+                f"{f' on node {args.decode_node!r}' if args.decode_node else ''}."
+            )
+            return 1
+        prefill_p = prefill_placements[0]
+        decode_p = decode_placements[0]
+
+    prefill_node_names = _placement_node_friendly_names(prefill_p, id_to_friendly)
+    decode_node_names = _placement_node_friendly_names(decode_p, id_to_friendly)
+    _ = unwrap_instance
+
+    prefill_instance = prefill_p["instance"]
+    decode_instance = decode_p["instance"]
+    prefill_id = instance_id_from_instance(prefill_instance)
+    decode_id = instance_id_from_instance(decode_instance)
+    prefill_meta = str(prefill_p.get("instance_meta", ""))
+    decode_meta = str(decode_p.get("instance_meta", ""))
+    prefill_nodes = nodes_used_in_instance(prefill_instance)
+    decode_nodes = nodes_used_in_instance(decode_instance)
+
+    logger.info("=" * 80)
+    logger.info(
+        f"PREFILL: {prefill_meta} / nodes={prefill_nodes} ({','.join(prefill_node_names)}) "
+        f"/ {prefill_short_id} ({prefill_full_id}) / instance_id={prefill_id}"
+    )
+    logger.info(
+        f"DECODE:  {decode_meta} / nodes={decode_nodes} ({','.join(decode_node_names)}) "
+        f"/ {decode_short_id} ({decode_full_id}) / instance_id={decode_id}"
+    )
+
+    if args.dry_run:
+        return 0
+
+    settle_deadline = (
+        time.monotonic() + args.settle_timeout if args.settle_timeout > 0 else None
+    )
+
+    logger.info("Planning phase: prefill...")
+    run_planning_phase(
+        client,
+        prefill_full_id,
+        prefill_p,
+        args.danger_delete_downloads,
+        args.timeout,
+        settle_deadline,
+    )
+    logger.info("Planning phase: decode...")
+    run_planning_phase(
+        client,
+        decode_full_id,
+        decode_p,
+        args.danger_delete_downloads,
+        args.timeout,
+        settle_deadline,
+    )
+
+    if use_combinations:
+        pp_tg_pairs = list(itertools.product(pp_list, tg_list))
+    else:
+        pp_tg_pairs = list(zip(pp_list, tg_list, strict=True))
+
+    common_meta = {
+        "decode_model_short_id": decode_short_id,
+        "decode_model_id": decode_full_id,
+        "prefill_model_short_id": prefill_short_id,
+        "prefill_model_id": prefill_full_id,
+        "prefill_instance_id": prefill_id,
+        "prefill_instance_meta": prefill_meta,
+        "prefill_nodes": prefill_nodes,
+        "decode_instance_id": decode_id,
+        "decode_instance_meta": decode_meta,
+        "decode_nodes": decode_nodes,
+    }
+
+    all_rows: list[dict[str, Any]] = []
+    disagg_rows: list[dict[str, Any]] = []
+    decode_alone_rows: list[dict[str, Any]] = []
+    prefill_alone_rows: list[dict[str, Any]] = []
+    link_id = ""
+    prefill_alive = False
+    decode_alive = False
+    try:
+        logger.info("Creating prefill instance...")
+        client.request_json("POST", "/instance", body={"instance": prefill_instance})
+        wait_for_instance_ready(client, prefill_id)
+        prefill_alive = True
+        logger.info("Prefill instance ready")
+
+        if args.compare_baseline:
+            time.sleep(2)
+            prefill_alone_rows = _run_phase(
+                client=client,
+                label="prefill_alone",
+                pp_tg_pairs=pp_tg_pairs,
+                model_id=prefill_full_id,
+                prompt_sizer=prefill_prompt_sizer,
+                warmup=args.warmup,
+                repeat=args.repeat,
+                common_meta=common_meta,
+            )
+            all_rows.extend(prefill_alone_rows)
+
+        logger.info("Creating decode instance...")
+        client.request_json("POST", "/instance", body={"instance": decode_instance})
+        wait_for_instance_ready(client, decode_id)
+        decode_alive = True
+        logger.info("Decode instance ready")
+
+        logger.info("Linking instances (prefill → decode)...")
+        _create_instance_link(client, prefill_id, decode_id)
+        time.sleep(1)
+        links = _list_instance_links(client)
+        if not links:
+            logger.error("Link did not appear in state.")
+            return 1
+        link_id = str(links[-1].get("linkId") or links[-1].get("link_id") or "")
+        logger.info(f"Link created: {link_id}")
+        time.sleep(2)
+
+        disagg_rows = _run_phase(
+            client=client,
+            label="disaggregated",
+            pp_tg_pairs=pp_tg_pairs,
+            model_id=decode_full_id,
+            prompt_sizer=decode_prompt_sizer,
+            warmup=args.warmup,
+            repeat=args.repeat,
+            common_meta=common_meta,
+        )
+        all_rows.extend(disagg_rows)
+
+        if args.compare_baseline:
+            logger.info("Removing link and prefill instance to isolate decode_alone.")
+            with contextlib.suppress(ExoHttpError):
+                if link_id:
+                    _delete_instance_link(client, link_id)
+                    link_id = ""
+            with contextlib.suppress(ExoHttpError):
+                client.request_json("DELETE", f"/instance/{prefill_id}")
+            wait_for_instance_gone(client, prefill_id)
+            prefill_alive = False
+            time.sleep(2)
+
+            decode_alone_rows = _run_phase(
+                client=client,
+                label="decode_alone",
+                pp_tg_pairs=pp_tg_pairs,
+                model_id=decode_full_id,
+                prompt_sizer=decode_prompt_sizer,
+                warmup=args.warmup,
+                repeat=args.repeat,
+                common_meta=common_meta,
+            )
+            all_rows.extend(decode_alone_rows)
+
+            _print_diff(disagg_rows, decode_alone_rows, prefill_alone_rows)
+    finally:
+        with contextlib.suppress(ExoHttpError):
+            if link_id:
+                _delete_instance_link(client, link_id)
+        if decode_alive:
+            with contextlib.suppress(ExoHttpError):
+                client.request_json("DELETE", f"/instance/{decode_id}")
+            wait_for_instance_gone(client, decode_id)
+        if prefill_alive:
+            with contextlib.suppress(ExoHttpError):
+                client.request_json("DELETE", f"/instance/{prefill_id}")
+            wait_for_instance_gone(client, prefill_id)
+        logger.debug("Deleted both instances")
+
+    if args.stdout:
+        json.dump(all_rows, sys.stdout, indent=2, ensure_ascii=False)
+    elif args.json_out:
+        with open(args.json_out, "w", encoding="utf-8") as f:
+            json.dump(all_rows, f, indent=2, ensure_ascii=False)
+        logger.debug(f"\nWrote results JSON: {args.json_out}")
+
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())
@@ -11,7 +11,8 @@
 				"highlight.js": "^11.11.1",
 				"katex": "^0.16.27",
 				"marked": "^17.0.1",
-				"mode-watcher": "^1.1.0"
+				"mode-watcher": "^1.1.0",
+				"pdfjs-dist": "^5.6.205"
 			},
 			"devDependencies": {
 				"@sveltejs/adapter-static": "^3.0.10",
@@ -518,6 +519,256 @@
 				"@jridgewell/sourcemap-codec": "^1.4.14"
 			}
 		},
+		"node_modules/@napi-rs/canvas": {
+			"version": "0.1.97",
+			"resolved": "https://registry.npmjs.org/@napi-rs/canvas/-/canvas-0.1.97.tgz",
+			"integrity": "sha512-8cFniXvrIEnVwuNSRCW9wirRZbHvrD3JVujdS2P5n5xiJZNZMOZcfOvJ1pb66c7jXMKHHglJEDVJGbm8XWFcXQ==",
+			"license": "MIT",
+			"optional": true,
+			"workspaces": [
+				"e2e/*"
+			],
+			"engines": {
+				"node": ">= 10"
+			},
+			"funding": {
+				"type": "github",
+				"url": "https://github.com/sponsors/Brooooooklyn"
+			},
+			"optionalDependencies": {
+				"@napi-rs/canvas-android-arm64": "0.1.97",
+				"@napi-rs/canvas-darwin-arm64": "0.1.97",
+				"@napi-rs/canvas-darwin-x64": "0.1.97",
+				"@napi-rs/canvas-linux-arm-gnueabihf": "0.1.97",
+				"@napi-rs/canvas-linux-arm64-gnu": "0.1.97",
+				"@napi-rs/canvas-linux-arm64-musl": "0.1.97",
+				"@napi-rs/canvas-linux-riscv64-gnu": "0.1.97",
+				"@napi-rs/canvas-linux-x64-gnu": "0.1.97",
+				"@napi-rs/canvas-linux-x64-musl": "0.1.97",
+				"@napi-rs/canvas-win32-arm64-msvc": "0.1.97",
+				"@napi-rs/canvas-win32-x64-msvc": "0.1.97"
+			}
+		},
+		"node_modules/@napi-rs/canvas-android-arm64": {
+			"version": "0.1.97",
+			"resolved": "https://registry.npmjs.org/@napi-rs/canvas-android-arm64/-/canvas-android-arm64-0.1.97.tgz",
+			"integrity": "sha512-V1c/WVw+NzH8vk7ZK/O8/nyBSCQimU8sfMsB/9qeSvdkGKNU7+mxy/bIF0gTgeBFmHpj30S4E9WHMSrxXGQuVQ==",
+			"cpu": [
+				"arm64"
+			],
+			"license": "MIT",
+			"optional": true,
+			"os": [
+				"android"
+			],
+			"engines": {
+				"node": ">= 10"
+			},
+			"funding": {
+				"type": "github",
+				"url": "https://github.com/sponsors/Brooooooklyn"
+			}
+		},
+		"node_modules/@napi-rs/canvas-darwin-arm64": {
+			"version": "0.1.97",
+			"resolved": "https://registry.npmjs.org/@napi-rs/canvas-darwin-arm64/-/canvas-darwin-arm64-0.1.97.tgz",
+			"integrity": "sha512-ok+SCEF4YejcxuJ9Rm+WWunHHpf2HmiPxfz6z1a/NFQECGXtsY7A4B8XocK1LmT1D7P174MzwPF9Wy3AUAwEPw==",
+			"cpu": [
+				"arm64"
+			],
+			"license": "MIT",
+			"optional": true,
+			"os": [
+				"darwin"
+			],
+			"engines": {
+				"node": ">= 10"
+			},
+			"funding": {
+				"type": "github",
+				"url": "https://github.com/sponsors/Brooooooklyn"
+			}
+		},
+		"node_modules/@napi-rs/canvas-darwin-x64": {
+			"version": "0.1.97",
+			"resolved": "https://registry.npmjs.org/@napi-rs/canvas-darwin-x64/-/canvas-darwin-x64-0.1.97.tgz",
+			"integrity": "sha512-PUP6e6/UGlclUvAQNnuXCcnkpdUou6VYZfQOQxExLp86epOylmiwLkqXIvpFmjoTEDmPmXrI+coL/9EFU1gKPA==",
+			"cpu": [
+				"x64"
+			],
+			"license": "MIT",
+			"optional": true,
+			"os": [
+				"darwin"
+			],
+			"engines": {
+				"node": ">= 10"
+			},
+			"funding": {
+				"type": "github",
+				"url": "https://github.com/sponsors/Brooooooklyn"
+			}
+		},
+		"node_modules/@napi-rs/canvas-linux-arm-gnueabihf": {
+			"version": "0.1.97",
+			"resolved": "https://registry.npmjs.org/@napi-rs/canvas-linux-arm-gnueabihf/-/canvas-linux-arm-gnueabihf-0.1.97.tgz",
+			"integrity": "sha512-XyXH2L/cic8eTNtbrXCcvqHtMX/nEOxN18+7rMrAM2XtLYC/EB5s0wnO1FsLMWmK+04ZSLN9FBGipo7kpIkcOw==",
+			"cpu": [
+				"arm"
+			],
+			"license": "MIT",
+			"optional": true,
+			"os": [
+				"linux"
+			],
+			"engines": {
+				"node": ">= 10"
+			},
+			"funding": {
+				"type": "github",
+				"url": "https://github.com/sponsors/Brooooooklyn"
+			}
+		},
+		"node_modules/@napi-rs/canvas-linux-arm64-gnu": {
+			"version": "0.1.97",
+			"resolved": "https://registry.npmjs.org/@napi-rs/canvas-linux-arm64-gnu/-/canvas-linux-arm64-gnu-0.1.97.tgz",
+			"integrity": "sha512-Kuq/M3djq0K8ktgz6nPlK7Ne5d4uWeDxPpyKWOjWDK2RIOhHVtLtyLiJw2fuldw7Vn4mhw05EZXCEr4Q76rs9w==",
+			"cpu": [
+				"arm64"
+			],
+			"license": "MIT",
+			"optional": true,
+			"os": [
+				"linux"
+			],
+			"engines": {
+				"node": ">= 10"
+			},
+			"funding": {
+				"type": "github",
+				"url": "https://github.com/sponsors/Brooooooklyn"
+			}
+		},
+		"node_modules/@napi-rs/canvas-linux-arm64-musl": {
+			"version": "0.1.97",
+			"resolved": "https://registry.npmjs.org/@napi-rs/canvas-linux-arm64-musl/-/canvas-linux-arm64-musl-0.1.97.tgz",
+			"integrity": "sha512-kKmSkQVnWeqg7qdsiXvYxKhAFuHz3tkBjW/zyQv5YKUPhotpaVhpBGv5LqCngzyuRV85SXoe+OFj+Tv0a0QXkQ==",
+			"cpu": [
+				"arm64"
+			],
+			"license": "MIT",
+			"optional": true,
+			"os": [
+				"linux"
+			],
+			"engines": {
+				"node": ">= 10"
+			},
+			"funding": {
+				"type": "github",
+				"url": "https://github.com/sponsors/Brooooooklyn"
+			}
+		},
+		"node_modules/@napi-rs/canvas-linux-riscv64-gnu": {
+			"version": "0.1.97",
+			"resolved": "https://registry.npmjs.org/@napi-rs/canvas-linux-riscv64-gnu/-/canvas-linux-riscv64-gnu-0.1.97.tgz",
+			"integrity": "sha512-Jc7I3A51jnEOIAXeLsN/M/+Z28LUeakcsXs07FLq9prXc0eYOtVwsDEv913Gr+06IRo34gJJVgT0TXvmz+N2VA==",
+			"cpu": [
+				"riscv64"
+			],
+			"license": "MIT",
+			"optional": true,
+			"os": [
+				"linux"
+			],
+			"engines": {
+				"node": ">= 10"
+			},
+			"funding": {
+				"type": "github",
+				"url": "https://github.com/sponsors/Brooooooklyn"
+			}
+		},
+		"node_modules/@napi-rs/canvas-linux-x64-gnu": {
+			"version": "0.1.97",
+			"resolved": "https://registry.npmjs.org/@napi-rs/canvas-linux-x64-gnu/-/canvas-linux-x64-gnu-0.1.97.tgz",
+			"integrity": "sha512-iDUBe7AilfuBSRbSa8/IGX38Mf+iCSBqoVKLSQ5XaY2JLOaqz1TVyPFEyIck7wT6mRQhQt5sN6ogfjIDfi74tg==",
+			"cpu": [
+				"x64"
+			],
+			"license": "MIT",
+			"optional": true,
+			"os": [
+				"linux"
+			],
+			"engines": {
+				"node": ">= 10"
+			},
+			"funding": {
+				"type": "github",
+				"url": "https://github.com/sponsors/Brooooooklyn"
+			}
+		},
+		"node_modules/@napi-rs/canvas-linux-x64-musl": {
+			"version": "0.1.97",
+			"resolved": "https://registry.npmjs.org/@napi-rs/canvas-linux-x64-musl/-/canvas-linux-x64-musl-0.1.97.tgz",
+			"integrity": "sha512-AKLFd/v0Z5fvgqBDqhvqtAdx+fHMJ5t9JcUNKq4FIZ5WH+iegGm8HPdj00NFlCSnm83Fp3Ln8I2f7uq1aIiWaA==",
+			"cpu": [
+				"x64"
+			],
+			"license": "MIT",
+			"optional": true,
+			"os": [
+				"linux"
+			],
+			"engines": {
+				"node": ">= 10"
+			},
+			"funding": {
+				"type": "github",
+				"url": "https://github.com/sponsors/Brooooooklyn"
+			}
+		},
+		"node_modules/@napi-rs/canvas-win32-arm64-msvc": {
+			"version": "0.1.97",
+			"resolved": "https://registry.npmjs.org/@napi-rs/canvas-win32-arm64-msvc/-/canvas-win32-arm64-msvc-0.1.97.tgz",
+			"integrity": "sha512-u883Yr6A6fO7Vpsy9YE4FVCIxzzo5sO+7pIUjjoDLjS3vQaNMkVzx5bdIpEL+ob+gU88WDK4VcxYMZ6nmnoX9A==",
+			"cpu": [
+				"arm64"
+			],
+			"license": "MIT",
+			"optional": true,
+			"os": [
+				"win32"
+			],
+			"engines": {
+				"node": ">= 10"
+			},
+			"funding": {
+				"type": "github",
+				"url": "https://github.com/sponsors/Brooooooklyn"
+			}
+		},
+		"node_modules/@napi-rs/canvas-win32-x64-msvc": {
+			"version": "0.1.97",
+			"resolved": "https://registry.npmjs.org/@napi-rs/canvas-win32-x64-msvc/-/canvas-win32-x64-msvc-0.1.97.tgz",
+			"integrity": "sha512-sWtD2EE3fV0IzN+iiQUqr/Q1SwqWhs2O1FKItFlxtdDkikpEj5g7DKQpY3x55H/MAOnL8iomnlk3mcEeGiUMoQ==",
+			"cpu": [
+				"x64"
+			],
+			"license": "MIT",
+			"optional": true,
+			"os": [
+				"win32"
+			],
+			"engines": {
+				"node": ">= 10"
+			},
+			"funding": {
+				"type": "github",
+				"url": "https://github.com/sponsors/Brooooooklyn"
+			}
+		},
 		"node_modules/@polka/url": {
 			"version": "1.0.0-next.29",
 			"resolved": "https://registry.npmjs.org/@polka/url/-/url-1.0.0-next.29.tgz",
@@ -2635,6 +2886,26 @@
 				"node": "^10 || ^12 || ^13.7 || ^14 || >=15.0.1"
 			}
 		},
+		"node_modules/node-readable-to-web-readable-stream": {
+			"version": "0.4.2",
+			"resolved": "https://registry.npmjs.org/node-readable-to-web-readable-stream/-/node-readable-to-web-readable-stream-0.4.2.tgz",
+			"integrity": "sha512-/cMZNI34v//jUTrI+UIo4ieHAB5EZRY/+7OmXZgBxaWBMcW2tGdceIw06RFxWxrKZ5Jp3sI2i5TsRo+CBhtVLQ==",
+			"license": "MIT",
+			"optional": true
+		},
+		"node_modules/pdfjs-dist": {
+			"version": "5.6.205",
+			"resolved": "https://registry.npmjs.org/pdfjs-dist/-/pdfjs-dist-5.6.205.tgz",
+			"integrity": "sha512-tlUj+2IDa7G1SbvBNN74UHRLJybZDWYom+k6p5KIZl7huBvsA4APi6mKL+zCxd3tLjN5hOOEE9Tv7VdzO88pfg==",
+			"license": "Apache-2.0",
+			"engines": {
+				"node": ">=20.19.0 || >=22.13.0 || >=24"
+			},
+			"optionalDependencies": {
+				"@napi-rs/canvas": "^0.1.96",
+				"node-readable-to-web-readable-stream": "^0.4.2"
+			}
+		},
 		"node_modules/picocolors": {
 			"version": "1.1.1",
 			"resolved": "https://registry.npmjs.org/picocolors/-/picocolors-1.1.1.tgz",
@@ -31,6 +31,7 @@
 		"highlight.js": "^11.11.1",
 		"katex": "^0.16.27",
 		"marked": "^17.0.1",
-		"mode-watcher": "^1.1.0"
+		"mode-watcher": "^1.1.0",
+		"pdfjs-dist": "^5.6.205"
 	}
 }
@@ -139,6 +139,8 @@
        return "🖼";
      case "text":
        return "📄";
+      case "pdf":
+        return "📑";
      default:
        return "📎";
    }
@@ -94,6 +94,12 @@
      d="M8.948 8.798v-1.43a6.7 6.7 0 0 1 .424-.018c3.922-.124 6.493 3.374 6.493 3.374s-2.774 3.851-5.75 3.851c-.398 0-.787-.062-1.158-.185v-4.346c1.528.185 1.837.857 2.747 2.385l2.04-1.714s-1.492-1.952-4-1.952a6.016 6.016 0 0 0-.796.035m0-4.735v2.138l.424-.027c5.45-.185 9.01 4.47 9.01 4.47s-4.08 4.964-8.33 4.964c-.37 0-.733-.035-1.095-.097v1.325c.3.035.61.062.91.062 3.957 0 6.82-2.023 9.593-4.408.459.371 2.34 1.263 2.73 1.652-2.633 2.208-8.772 3.984-12.253 3.984-.335 0-.653-.018-.971-.053v1.864H24V4.063zm0 10.326v1.131c-3.657-.654-4.673-4.46-4.673-4.46s1.758-1.944 4.673-2.262v1.237H8.94c-1.528-.186-2.73 1.245-2.73 1.245s.68 2.412 2.739 3.11M2.456 10.9s2.164-3.197 6.5-3.533V6.201C4.153 6.59 0 10.653 0 10.653s2.35 6.802 8.948 7.42v-1.237c-4.84-.6-6.492-5.936-6.492-5.936z"
    />
  </svg>
+{:else if family === "gemma"}
+  <svg class="w-6 h-6 {className}" viewBox="0 0 24 24" fill="currentColor">
+    <path
+      d="M12.48 10.92v3.28h7.84c-.24 1.84-.853 3.187-1.787 4.133-1.147 1.147-2.933 2.4-6.053 2.4-4.827 0-8.6-3.893-8.6-8.72s3.773-8.72 8.6-8.72c2.6 0 4.507 1.027 5.907 2.347l2.307-2.307C18.747 1.44 16.133 0 12.48 0 5.867 0 .307 5.387.307 12s5.56 12 12.173 12c3.573 0 6.267-1.173 8.373-3.36 2.16-2.16 2.84-5.213 2.84-7.667 0-.76-.053-1.467-.173-2.053H12.48z"
+    />
+  </svg>
 {:else}
  <svg class="w-6 h-6 {className}" viewBox="0 0 24 24" fill="currentColor">
    <path
@@ -32,6 +32,7 @@
    flux: "FLUX",
    "qwen-image": "Qwen Img",
    nemotron: "NVIDIA",
+    gemma: "Google",
  };

  function getFamilyName(family: string): string {
@@ -42,31 +43,20 @@
 </script>

 <div
-  class="flex flex-col gap-1 py-2 px-1 border-r border-exo-yellow/10 bg-exo-medium-gray/30 min-w-[72px] sm:min-w-[64px] overflow-y-auto scrollbar-hide"
+  class="flex flex-col gap-1 py-2 px-1 border-r border-exo-yellow/10 bg-exo-medium-gray/30 min-w-[80px] sm:min-w-[72px] overflow-y-auto scrollbar-hide"
 >
  <!-- All models (no filter) -->
  <button
    type="button"
    onclick={() => onSelect(null)}
-    class="group flex flex-col items-center justify-center p-2 sm:p-2 rounded transition-all duration-200 cursor-pointer min-h-[44px] sm:min-h-0 {selectedFamily ===
+    class="group flex items-center justify-center px-3 py-2.5 rounded transition-all duration-200 cursor-pointer min-h-[44px] sm:min-h-0 {selectedFamily ===
    null
      ? 'bg-exo-yellow/20 border-l-2 border-exo-yellow'
      : 'hover:bg-white/5 border-l-2 border-transparent'}"
    title="All models"
  >
-    <svg
-      class="w-5 h-5 {selectedFamily === null
-        ? 'text-exo-yellow'
-        : 'text-white/50 group-hover:text-white/70'}"
-      viewBox="0 0 24 24"
-      fill="currentColor"
-    >
-      <path
-        d="M4 8h4V4H4v4zm6 12h4v-4h-4v4zm-6 0h4v-4H4v4zm0-6h4v-4H4v4zm6 0h4v-4h-4v4zm6-10v4h4V4h-4zm-6 4h4V4h-4v4zm6 6h4v-4h-4v4zm0 6h4v-4h-4v4z"
-      />
-    </svg>
    <span
-      class="text-[9px] font-mono mt-0.5 {selectedFamily === null
+      class="text-[12px] font-mono font-medium {selectedFamily === null
        ? 'text-exo-yellow'
        : 'text-white/40 group-hover:text-white/60'}">All</span
    >
@@ -90,7 +80,7 @@
          : "text-white/50 group-hover:text-amber-400/70"}
      />
      <span
-        class="text-[9px] font-mono mt-0.5 {selectedFamily === 'favorites'
+        class="text-[11px] font-mono mt-0.5 {selectedFamily === 'favorites'
          ? 'text-amber-400'
          : 'text-white/40 group-hover:text-white/60'}">Faves</span
      >
@@ -115,7 +105,7 @@
          : "text-white/50 group-hover:text-white/70"}
      />
      <span
-        class="text-[9px] font-mono mt-0.5 {selectedFamily === 'recents'
+        class="text-[11px] font-mono mt-0.5 {selectedFamily === 'recents'
          ? 'text-exo-yellow'
          : 'text-white/40 group-hover:text-white/60'}">Recent</span
      >
@@ -139,7 +129,7 @@
        : "text-white/50 group-hover:text-orange-400/70"}
    />
    <span
-      class="text-[9px] font-mono mt-0.5 {selectedFamily === 'huggingface'
+      class="text-[11px] font-mono mt-0.5 {selectedFamily === 'huggingface'
        ? 'text-orange-400'
        : 'text-white/40 group-hover:text-white/60'}">Hub</span
    >
@@ -165,7 +155,7 @@
          : "text-white/50 group-hover:text-white/70"}
      />
      <span
-        class="text-[9px] font-mono mt-0.5 truncate max-w-full {selectedFamily ===
+        class="text-[11px] font-mono mt-0.5 truncate max-w-full {selectedFamily ===
        family
          ? 'text-exo-yellow'
          : 'text-white/40 group-hover:text-white/60'}"
@@ -1,21 +1,41 @@
 <script lang="ts">
  import { browser } from "$app/environment";
+  import { featureFlags } from "$lib/stores/app.svelte";

-  export let showHome = true;
-  export let onHome: (() => void) | null = null;
-  export let showSidebarToggle = false;
-  export let sidebarVisible = true;
-  export let onToggleSidebar: (() => void) | null = null;
-  export let showMobileMenuToggle = false;
-  export let mobileMenuOpen = false;
-  export let onToggleMobileMenu: (() => void) | null = null;
-  export let showMobileRightToggle = false;
-  export let mobileRightOpen = false;
-  export let onToggleMobileRight: (() => void) | null = null;
-  export let downloadProgress: {
-    count: number;
-    percentage: number;
-  } | null = null;
+  const showAdvanced = $derived(featureFlags()["disaggregation"] === true);
+
+  interface Props {
+    showHome?: boolean;
+    onHome?: (() => void) | null;
+    showSidebarToggle?: boolean;
+    sidebarVisible?: boolean;
+    onToggleSidebar?: (() => void) | null;
+    showMobileMenuToggle?: boolean;
+    mobileMenuOpen?: boolean;
+    onToggleMobileMenu?: (() => void) | null;
+    showMobileRightToggle?: boolean;
+    mobileRightOpen?: boolean;
+    onToggleMobileRight?: (() => void) | null;
+    downloadProgress?: {
+      count: number;
+      percentage: number;
+    } | null;
+  }
+
+  let {
+    showHome = true,
+    onHome = null,
+    showSidebarToggle = false,
+    sidebarVisible = true,
+    onToggleSidebar = null,
+    showMobileMenuToggle = false,
+    mobileMenuOpen = false,
+    onToggleMobileMenu = null,
+    showMobileRightToggle = false,
+    mobileRightOpen = false,
+    onToggleMobileRight = null,
+    downloadProgress = null,
+  }: Props = $props();

  function handleHome(): void {
    if (onHome) {
@@ -259,5 +279,49 @@
      {/if}
      <span class="hidden sm:inline">Downloads</span>
    </a>
+    <a
+      href="/#/integrations"
+      class="text-xs md:text-sm text-white/70 hover:text-exo-yellow transition-colors tracking-wider uppercase flex items-center gap-1.5 md:gap-2 cursor-pointer"
+      title="Integration configs for external tools"
+    >
+      <svg
+        class="w-4 h-4"
+        viewBox="0 0 24 24"
+        fill="none"
+        stroke="currentColor"
+        stroke-width="2"
+        stroke-linecap="round"
+        stroke-linejoin="round"
+      >
+        <path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71" />
+        <path
+          d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"
+        />
+      </svg>
+      <span class="hidden sm:inline">Integrations</span>
+    </a>
+    {#if showAdvanced}
+      <a
+        href="/#/advanced"
+        class="text-xs md:text-sm text-white/70 hover:text-exo-yellow transition-colors tracking-wider uppercase flex items-center gap-1.5 md:gap-2 cursor-pointer"
+        title="Advanced cluster settings"
+      >
+        <svg
+          class="w-4 h-4"
+          viewBox="0 0 24 24"
+          fill="none"
+          stroke="currentColor"
+          stroke-width="2"
+          stroke-linecap="round"
+          stroke-linejoin="round"
+        >
+          <circle cx="12" cy="12" r="3" />
+          <path
+            d="M19.4 15a1.65 1.65 0 0 0 .33 1.82l.06.06a2 2 0 0 1 0 2.83 2 2 0 0 1-2.83 0l-.06-.06a1.65 1.65 0 0 0-1.82-.33 1.65 1.65 0 0 0-1 1.51V21a2 2 0 0 1-4 0v-.09A1.65 1.65 0 0 0 9 19.4a1.65 1.65 0 0 0-1.82.33l-.06.06a2 2 0 0 1-2.83 0 2 2 0 0 1 0-2.83l.06-.06a1.65 1.65 0 0 0 .33-1.82 1.65 1.65 0 0 0-1.51-1H3a2 2 0 0 1 0-4h.09A1.65 1.65 0 0 0 4.6 9a1.65 1.65 0 0 0-.33-1.82l-.06-.06a2 2 0 0 1 0-2.83 2 2 0 0 1 2.83 0l.06.06a1.65 1.65 0 0 0 1.82.33H9a1.65 1.65 0 0 0 1-1.51V3a2 2 0 0 1 4 0v.09a1.65 1.65 0 0 0 1 1.51 1.65 1.65 0 0 0 1.82-.33l.06-.06a2 2 0 0 1 2.83 0 2 2 0 0 1 0 2.83l-.06.06a1.65 1.65 0 0 0-.33 1.82V9a1.65 1.65 0 0 0 1.51 1H21a2 2 0 0 1 0 4h-.09a1.65 1.65 0 0 0-1.51 1z"
+          />
+        </svg>
+        <span class="hidden sm:inline">Advanced</span>
+      </a>
+    {/if}
  </nav>
 </header>
@@ -0,0 +1,62 @@
+<script lang="ts">
+  import { copyText } from "$lib/utils/clipboard";
+
+  interface Props {
+    title: string;
+    subtitle: string;
+    config: string;
+    description?: string;
+    language?: "json" | "bash";
+  }
+
+  let {
+    title,
+    subtitle,
+    config,
+    description = "",
+    language = "json",
+  }: Props = $props();
+
+  let copied = $state(false);
+  let failed = $state(false);
+
+  async function copyToClipboard() {
+    const ok = await copyText(config);
+    if (ok) {
+      copied = true;
+      setTimeout(() => (copied = false), 2000);
+    } else {
+      failed = true;
+      setTimeout(() => (failed = false), 2000);
+    }
+  }
+</script>
+
+<div
+  class="border border-exo-light-gray/20 rounded-lg bg-exo-medium-gray/20 overflow-hidden"
+>
+  <div class="flex items-center justify-between px-5 py-4">
+    <div>
+      <h3 class="text-white text-sm font-semibold tracking-wide">{title}</h3>
+      <p class="text-exo-light-gray/60 text-xs mt-0.5 font-mono">{subtitle}</p>
+    </div>
+    <button
+      onclick={copyToClipboard}
+      class="px-3 py-1.5 text-xs rounded border transition-all duration-200 cursor-pointer
+        {copied
+        ? 'border-green-500/50 text-green-400 bg-green-500/10'
+        : failed
+          ? 'border-red-500/50 text-red-400 bg-red-500/10'
+          : 'border-exo-light-gray/30 text-exo-light-gray hover:border-exo-yellow/50 hover:text-exo-yellow'}"
+    >
+      {copied ? "Copied!" : failed ? "Copy failed" : "Copy"}
+    </button>
+  </div>
+  {#if description}
+    <p class="text-exo-light-gray/70 text-xs px-5 pb-3">{description}</p>
+  {/if}
+  <div class="bg-black/30 border-t border-exo-light-gray/10">
+    <pre
+      class="text-xs text-exo-light-gray/90 font-mono p-4 overflow-x-auto whitespace-pre">{config}</pre>
+  </div>
+</div>
@@ -73,7 +73,7 @@

 <!-- svelte-ignore a11y_no_static_element_interactions -->
 <div
-  class="filter-popover absolute right-0 top-full mt-2 w-64 bg-exo-dark-gray border border-exo-yellow/10 rounded-lg shadow-xl z-10"
+  class="filter-popover absolute right-0 top-full mt-2 w-64 bg-exo-dark-gray border border-exo-yellow/10 rounded-lg shadow-xl z-20"
  transition:fly={{ y: -10, duration: 200, easing: cubicOut }}
  onclick={(e) => e.stopPropagation()}
  role="dialog"
@@ -457,6 +457,7 @@
      "deepseek",
      "gpt-oss",
      "llama",
+      "gemma",
      "flux",
      "qwen-image",
      "nemotron",
@@ -0,0 +1,565 @@
+<script lang="ts">
+  import { onMount, onDestroy } from "svelte";
+  import FamilyLogos from "$lib/components/FamilyLogos.svelte";
+  import {
+    instances,
+    instanceLinks,
+    nodeIdentities,
+    refreshState,
+    createInstanceLink,
+    updateInstanceLink,
+    deleteInstanceLink,
+    type Instance,
+  } from "$lib/stores/app.svelte";
+  import { deriveBaseModel, deriveFamily } from "$lib/utils/model_family";
+
+  type InstanceWrapper = {
+    MlxRingInstance?: Instance;
+    MlxJacclInstance?: Instance;
+    VllmInstance?: Instance;
+  };
+
+  let interval: ReturnType<typeof setInterval> | null = null;
+
+  onMount(() => {
+    refreshState();
+    interval = setInterval(refreshState, 3000);
+  });
+  onDestroy(() => {
+    if (interval) clearInterval(interval);
+  });
+
+  type InstanceRow = {
+    id: string;
+    modelId: string;
+    family: string;
+    baseModel: string;
+    nodeNames: string[];
+    nodeCount: number;
+  };
+
+  const instanceRows = $derived.by<InstanceRow[]>(() => {
+    const rows: InstanceRow[] = [];
+    const ids = nodeIdentities();
+    for (const [id, raw] of Object.entries(instances())) {
+      const wrapper = raw as InstanceWrapper;
+      const inst =
+        wrapper.MlxRingInstance ??
+        wrapper.MlxJacclInstance ??
+        wrapper.VllmInstance;
+      const modelId = inst?.shardAssignments?.modelId ?? "";
+      const nodeToRunner = inst?.shardAssignments?.nodeToRunner ?? {};
+      const nodeIds = Object.keys(nodeToRunner);
+      const nodeNames = nodeIds
+        .map((nodeId) => ids[nodeId]?.friendlyName ?? nodeId.slice(0, 6))
+        .filter((name) => !!name);
+      rows.push({
+        id,
+        modelId,
+        family: deriveFamily(modelId),
+        baseModel: deriveBaseModel(modelId),
+        nodeNames,
+        nodeCount: nodeIds.length,
+      });
+    }
+    rows.sort((a, b) => a.modelId.localeCompare(b.modelId));
+    return rows;
+  });
+
+  const instanceById = $derived(
+    Object.fromEntries(instanceRows.map((r) => [r.id, r])),
+  );
+
+  type LinkRow = {
+    linkId: string;
+    prefill: string[];
+    decode: string[];
+    families: string[];
+    multiNode: boolean;
+  };
+
+  const linkRows = $derived.by<LinkRow[]>(() => {
+    const rows: LinkRow[] = [];
+    for (const [, link] of Object.entries(instanceLinks())) {
+      const fams = new Set<string>();
+      let multiNode = false;
+      for (const id of [...link.prefillInstances, ...link.decodeInstances]) {
+        const r = instanceById[id];
+        if (r && r.baseModel) fams.add(r.baseModel.toLowerCase());
+        if (r && r.nodeCount > 1) multiNode = true;
+      }
+      rows.push({
+        linkId: link.linkId,
+        prefill: link.prefillInstances,
+        decode: link.decodeInstances,
+        families: Array.from(fams),
+        multiNode,
+      });
+    }
+    return rows;
+  });
+
+  let editingLinkId = $state<string | null>(null);
+  let editingPrefill = $state<Set<string>>(new Set());
+  let editingDecode = $state<Set<string>>(new Set());
+  let saving = $state(false);
+  let errorMessage = $state<string | null>(null);
+
+  function startCreate() {
+    editingLinkId = "new";
+    editingPrefill = new Set();
+    editingDecode = new Set();
+    errorMessage = null;
+  }
+
+  function startEdit(row: LinkRow) {
+    editingLinkId = row.linkId;
+    editingPrefill = new Set(row.prefill);
+    editingDecode = new Set(row.decode);
+    errorMessage = null;
+  }
+
+  function cancelEdit() {
+    editingLinkId = null;
+    editingPrefill = new Set();
+    editingDecode = new Set();
+    errorMessage = null;
+  }
+
+  type Role = "prefill" | "decode" | "none";
+
+  function roleOf(id: string): Role {
+    if (editingPrefill.has(id)) return "prefill";
+    if (editingDecode.has(id)) return "decode";
+    return "none";
+  }
+
+  function setRole(id: string, role: Role) {
+    const p = new Set(editingPrefill);
+    const d = new Set(editingDecode);
+    p.delete(id);
+    d.delete(id);
+    if (role === "prefill") p.add(id);
+    if (role === "decode") d.add(id);
+    editingPrefill = p;
+    editingDecode = d;
+  }
+
+  const editingFamilies = $derived.by<string[]>(() => {
+    const fams = new Set<string>();
+    for (const id of [...editingPrefill, ...editingDecode]) {
+      const r = instanceById[id];
+      if (r && r.baseModel) fams.add(r.baseModel.toLowerCase());
+    }
+    return Array.from(fams);
+  });
+
+  const editingMultiNode = $derived.by<string[]>(() => {
+    const names: string[] = [];
+    for (const id of [...editingPrefill, ...editingDecode]) {
+      const r = instanceById[id];
+      if (r && r.nodeCount > 1) {
+        names.push(r.baseModel || r.modelId);
+      }
+    }
+    return names;
+  });
+
+  const editingMismatch = $derived(editingFamilies.length > 1);
+  const canSave = $derived(
+    editingLinkId !== null &&
+      editingPrefill.size > 0 &&
+      editingDecode.size > 0 &&
+      !saving,
+  );
+
+  async function save() {
+    if (editingLinkId === null) return;
+    saving = true;
+    errorMessage = null;
+    try {
+      const prefill = Array.from(editingPrefill);
+      const decode = Array.from(editingDecode);
+      if (editingLinkId === "new") {
+        await createInstanceLink(prefill, decode);
+      } else {
+        await updateInstanceLink(editingLinkId, prefill, decode);
+      }
+      cancelEdit();
+      await refreshState();
+    } catch (err) {
+      errorMessage = err instanceof Error ? err.message : String(err);
+    } finally {
+      saving = false;
+    }
+  }
+
+  async function remove(linkId: string) {
+    if (!confirm("Remove this routing?")) return;
+    try {
+      await deleteInstanceLink(linkId);
+      if (editingLinkId === linkId) cancelEdit();
+      await refreshState();
+    } catch (err) {
+      errorMessage = err instanceof Error ? err.message : String(err);
+    }
+  }
+</script>
+
+<div class="font-mono text-foreground">
+  <div class="mb-6 space-y-4">
+    <details open class="group [&_summary::-webkit-details-marker]:hidden">
+      <summary
+        class="cursor-pointer list-none text-exo-yellow text-xs font-mono tracking-widest uppercase flex items-center gap-2 hover:opacity-80 transition-opacity"
+      >
+        <span
+          class="inline-block transition-transform group-open:rotate-90 text-exo-light-gray"
+          >▶</span
+        >
+        Prefill vs Decode
+      </summary>
+      <div class="mt-2 text-white/80 text-sm leading-relaxed">
+        Prefill is the compute-heavy pass that consumes the entire prompt and
+        builds a KV cache. Decode is the memory-bandwidth-bound loop that emits
+        tokens sequentially from that cache. The two phases have very different
+        bottlenecks, so running them on different hardware can be substantially
+        faster than doing both on one node.
+      </div>
+    </details>
+    <details class="group [&_summary::-webkit-details-marker]:hidden">
+      <summary
+        class="cursor-pointer list-none text-exo-yellow text-xs font-mono tracking-widest uppercase flex items-center gap-2 hover:opacity-80 transition-opacity"
+      >
+        <span
+          class="inline-block transition-transform group-open:rotate-90 text-exo-light-gray"
+          >▶</span
+        >
+        Linking Instances
+      </summary>
+      <div class="mt-2 text-white/80 text-sm leading-relaxed space-y-2">
+        <p>
+          A linked route here tells the cluster: when a request is sent to a
+          model in that cluster, the decode node (or the least active one if
+          there are multiple) will handle it. If it decides it must do a lot of
+          prefill not already cached in the prefix cache, it routes the request
+          to the prefill node over TCP IP. The prefill node streams the KV cache
+          back to the decode node which picks up from there.
+        </p>
+        <p>
+          Linked instances must be running the same model family — KV layouts
+          differ across architectures. More on the <a
+            class="text-exo-yellow underline underline-offset-2 hover:text-exo-yellow-darker transition-colors"
+            href="https://blog.exolabs.net/nvidia-dgx-spark/"
+            target="_blank"
+            rel="noreferrer noopener">blog</a
+          >.
+        </p>
+      </div>
+    </details>
+  </div>
+
+  {#if errorMessage}
+    <div
+      class="mb-4 px-4 py-3 bg-red-500/10 border border-red-500/40 text-red-300 text-sm"
+    >
+      {errorMessage}
+    </div>
+  {/if}
+
+  <section class="mt-12">
+    <h2
+      class="text-exo-yellow text-xs font-mono tracking-widest uppercase m-0 mb-3"
+    >
+      Existing routes
+    </h2>
+
+    {#if linkRows.length === 0}
+      {#if editingLinkId === null}
+        <div class="flex items-center justify-between">
+          <p class="text-exo-light-gray italic text-sm m-0">
+            No routes yet. Create one to enable remote prefill.
+          </p>
+          <button
+            class="px-3 py-1.5 text-xs font-mono tracking-wider uppercase bg-exo-yellow/15 border border-exo-yellow/50 text-exo-yellow hover:bg-exo-yellow/25 hover:border-exo-yellow/80 transition-colors"
+            onclick={startCreate}
+          >
+            + New route
+          </button>
+        </div>
+      {/if}
+    {:else}
+      {#if editingLinkId === null}
+        <div class="flex justify-end mb-3">
+          <button
+            class="px-3 py-1.5 text-xs font-mono tracking-wider uppercase bg-exo-yellow/15 border border-exo-yellow/50 text-exo-yellow hover:bg-exo-yellow/25 hover:border-exo-yellow/80 transition-colors"
+            onclick={startCreate}
+          >
+            + New route
+          </button>
+        </div>
+      {/if}
+      <div
+        class="bg-exo-dark-gray/60 border border-exo-medium-gray/40 flex flex-col"
+      >
+        {#each linkRows as row (row.linkId)}
+          {#if editingLinkId !== row.linkId}
+            <article
+              class="p-4 border-b border-exo-light-gray/25 last:border-b-0"
+            >
+              {#if row.multiNode}
+                <div
+                  class="mb-3 px-3 py-2 bg-red-500/10 border border-red-500/40 text-red-300 text-xs tracking-wide"
+                >
+                  ⚠ Multi-node instance detected. Remote prefill currently only
+                  works on single-node (rank-0) instances. This route will not
+                  function until that's supported.
+                </div>
+              {/if}
+              {#if row.families.length > 1}
+                <div
+                  class="mb-3 px-3 py-2 bg-amber-500/10 border border-amber-500/40 text-amber-300 text-xs tracking-wide"
+                >
+                  ⚠ Mixed model families: {row.families.join(", ")}
+                </div>
+              {/if}
+              <div
+                class="grid grid-cols-[1fr_auto_1fr_auto] items-center gap-x-3 gap-y-2"
+              >
+                <span
+                  class="inline-block justify-self-start text-[10px] font-mono tracking-widest uppercase px-2 py-0.5 bg-exo-yellow/15 border border-exo-yellow/40 text-exo-yellow"
+                  >Prefill</span
+                >
+                <span></span>
+                <span
+                  class="inline-block justify-self-start text-[10px] font-mono tracking-widest uppercase px-2 py-0.5 bg-exo-medium-gray/40 border border-exo-medium-gray/60 text-foreground"
+                  >Decode</span
+                >
+                <span></span>
+                <div class="min-w-0">
+                  <ul class="list-none p-0 m-0 flex flex-col gap-2">
+                    {#each row.prefill as id (id)}
+                      {@const r = instanceById[id]}
+                      {#if r}
+                        <li
+                          class="flex items-center gap-2 px-2.5 py-2 bg-exo-medium-gray/20 border border-exo-medium-gray/40"
+                        >
+                          <FamilyLogos family={r.family} />
+                          <div class="min-w-0 flex-1">
+                            <div
+                              class="text-exo-yellow text-xs font-mono truncate"
+                            >
+                              {r.baseModel || r.modelId}
+                            </div>
+                            <div
+                              class="text-exo-light-gray text-[11px] truncate"
+                            >
+                              {r.nodeNames.join(", ") || "?"}{r.nodeCount > 1
+                                ? ` (${r.nodeCount} nodes)`
+                                : ""}
+                            </div>
+                            <div
+                              class="text-exo-light-gray/40 text-[10px] font-mono truncate"
+                              title={r.id}
+                            >
+                              {r.id.slice(0, 8)}
+                            </div>
+                          </div>
+                        </li>
+                      {/if}
+                    {/each}
+                  </ul>
+                </div>
+                <div class="text-exo-yellow/60 text-xl px-2" aria-hidden="true">
+                  →
+                </div>
+                <div class="min-w-0">
+                  <ul class="list-none p-0 m-0 flex flex-col gap-2">
+                    {#each row.decode as id (id)}
+                      {@const r = instanceById[id]}
+                      {#if r}
+                        <li
+                          class="flex items-center gap-2 px-2.5 py-2 bg-exo-medium-gray/20 border border-exo-medium-gray/40"
+                        >
+                          <FamilyLogos family={r.family} />
+                          <div class="min-w-0 flex-1">
+                            <div
+                              class="text-exo-yellow text-xs font-mono truncate"
+                            >
+                              {r.baseModel || r.modelId}
+                            </div>
+                            <div
+                              class="text-exo-light-gray text-[11px] truncate"
+                            >
+                              {r.nodeNames.join(", ") || "?"}{r.nodeCount > 1
+                                ? ` (${r.nodeCount} nodes)`
+                                : ""}
+                            </div>
+                            <div
+                              class="text-exo-light-gray/40 text-[10px] font-mono truncate"
+                              title={r.id}
+                            >
+                              {r.id.slice(0, 8)}
+                            </div>
+                          </div>
+                        </li>
+                      {/if}
+                    {/each}
+                  </ul>
+                </div>
+                <div class="flex gap-2 pl-3">
+                  <button
+                    class="px-2 py-0.5 text-[11px] font-mono tracking-wider uppercase bg-exo-medium-gray/30 border border-exo-medium-gray/60 rounded text-foreground hover:border-exo-yellow/60 hover:text-exo-yellow disabled:opacity-40 disabled:cursor-not-allowed transition-colors"
+                    onclick={() => startEdit(row)}
+                    disabled={editingLinkId !== null}
+                  >
+                    Edit
+                  </button>
+                  <button
+                    class="px-2 py-0.5 text-[11px] font-mono tracking-wider uppercase bg-red-500/15 border border-red-500/40 rounded text-red-300 hover:bg-red-500/25 transition-colors"
+                    onclick={() => remove(row.linkId)}
+                  >
+                    Remove
+                  </button>
+                </div>
+              </div>
+            </article>
+          {/if}
+        {/each}
+      </div>
+    {/if}
+  </section>
+
+  {#if editingLinkId !== null && instanceRows.length === 0}
+    <section
+      class="mt-6 bg-exo-dark-gray/60 border border-exo-yellow/30 px-4 py-2.5 flex items-center justify-between gap-3"
+    >
+      <span class="text-exo-light-gray italic text-sm font-mono"
+        >No instances available.</span
+      >
+      <button
+        class="px-3 py-1 text-xs font-mono tracking-wider uppercase bg-exo-medium-gray/30 border border-exo-medium-gray/60 rounded text-foreground hover:border-exo-yellow/60 transition-colors"
+        onclick={cancelEdit}
+      >
+        Cancel
+      </button>
+    </section>
+  {:else if editingLinkId !== null}
+    <section class="mt-6 bg-exo-dark-gray/60 border border-exo-yellow/30 p-5">
+      <h2
+        class="text-exo-yellow text-xs font-mono tracking-widest uppercase m-0 mb-3"
+      >
+        {editingLinkId === "new" ? "New route" : "Edit route"}
+      </h2>
+
+      {#if editingMismatch}
+        <div
+          class="mb-3 px-3 py-2 bg-amber-500/10 border border-amber-500/40 text-amber-300 text-xs tracking-wide"
+        >
+          ⚠ Selected instances span multiple model families: <strong
+            >{editingFamilies.join(", ")}</strong
+          >. Linking across families produces a corrupt KV cache.
+        </div>
+      {/if}
+
+      {#if editingMultiNode.length > 0}
+        <div
+          class="mb-3 px-3 py-2 bg-red-500/10 border border-red-500/40 text-red-300 text-xs tracking-wide"
+        >
+          ⚠ Multi-node instance(s) selected: <strong
+            >{editingMultiNode.join(", ")}</strong
+          >. Remote prefill currently only works on single-node instances. This
+          route will not function until multi-node support lands.
+        </div>
+      {/if}
+
+      <p class="text-exo-light-gray text-xs mb-4">
+        Pick a role for each instance:
+        <span class="text-exo-yellow">Prefill</span>
+        serves KV cache,
+        <span class="text-foreground">Decode</span> consumes it.
+      </p>
+      <div
+        class="grid gap-2.5"
+        style="grid-template-columns: repeat(auto-fill, minmax(360px, 1fr));"
+      >
+        {#each instanceRows as row (row.id)}
+          {@const role = roleOf(row.id)}
+          <div
+            class="border p-3 flex flex-col gap-2.5 transition-colors {role ===
+            'prefill'
+              ? 'border-exo-yellow/60 bg-exo-dark-gray/60'
+              : role === 'decode'
+                ? 'border-exo-light-gray/60 bg-exo-dark-gray/60'
+                : 'border-exo-medium-gray/40 bg-exo-dark-gray/40'}"
+          >
+            <div class="flex items-center gap-2">
+              <FamilyLogos family={row.family} />
+              <div class="min-w-0 flex-1">
+                <div class="text-exo-yellow text-xs font-mono truncate">
+                  {row.baseModel || row.modelId}
+                </div>
+                <div class="text-exo-light-gray text-[11px] truncate">
+                  {row.nodeNames.join(", ") || "?"}{row.nodeCount > 1
+                    ? ` (${row.nodeCount} nodes)`
+                    : ""}
+                </div>
+                <div
+                  class="text-exo-light-gray/40 text-[10px] font-mono truncate"
+                  title={row.id}
+                >
+                  {row.id.slice(0, 8)}
+                </div>
+              </div>
+              {#if row.nodeCount > 1}
+                <span
+                  class="text-[9px] font-mono tracking-widest uppercase px-1.5 py-0.5 bg-red-500/15 border border-red-500/40 text-red-300"
+                  title="Multi-node instances are not supported by remote prefill yet."
+                  >Unsupported</span
+                >
+              {/if}
+            </div>
+            <div
+              class="flex rounded-md overflow-hidden border border-exo-light-gray/40 divide-x divide-exo-light-gray/40"
+            >
+              <button
+                class="flex-1 px-2 py-1 text-[11px] font-mono tracking-wider uppercase transition-colors {role ===
+                'prefill'
+                  ? 'bg-exo-yellow/20 text-exo-yellow'
+                  : 'bg-transparent text-white/80 hover:text-exo-yellow'}"
+                onclick={() =>
+                  setRole(row.id, role === "prefill" ? "none" : "prefill")}
+                >Prefill</button
+              >
+              <button
+                class="flex-1 px-2 py-1 text-[11px] font-mono tracking-wider uppercase transition-colors {role ===
+                'decode'
+                  ? 'bg-exo-medium-gray/50 text-foreground'
+                  : 'bg-transparent text-white/80 hover:text-foreground'}"
+                onclick={() =>
+                  setRole(row.id, role === "decode" ? "none" : "decode")}
+                >Decode</button
+              >
+            </div>
+          </div>
+        {/each}
+      </div>
+
+      <div class="flex gap-2 mt-5 justify-end">
+        <button
+          class="px-3 py-1.5 text-xs font-mono tracking-wider uppercase bg-exo-yellow/15 border border-exo-yellow/50 text-exo-yellow hover:bg-exo-yellow/25 hover:border-exo-yellow/80 disabled:opacity-40 disabled:cursor-not-allowed transition-colors"
+          onclick={save}
+          disabled={!canSave}
+        >
+          {saving ? "Saving..." : "Save route"}
+        </button>
+        <button
+          class="px-3 py-1.5 text-xs font-mono tracking-wider uppercase bg-exo-medium-gray/30 border border-exo-medium-gray/60 text-foreground hover:border-exo-yellow/60 disabled:opacity-40 disabled:cursor-not-allowed transition-colors"
+          onclick={cancelEdit}
+          disabled={saving}
+        >
+          Cancel
+        </button>
+      </div>
+    </section>
+  {/if}
+</div>
@@ -13,3 +13,4 @@ export { default as ModelFilterPopover } from "./ModelFilterPopover.svelte";
 export { default as ModelPickerGroup } from "./ModelPickerGroup.svelte";
 export { default as ModelPickerModal } from "./ModelPickerModal.svelte";
 export { default as ChatModelSelector } from "./ChatModelSelector.svelte";
+export { default as IntegrationCard } from "./IntegrationCard.svelte";
@@ -74,6 +74,12 @@ export interface Instance {
  };
 }

+export interface RawInstanceLink {
+  linkId: string;
+  prefillInstances: string[];
+  decodeInstances: string[];
+}
+
 // Granular node state types from the new state structure
 interface RawNodeIdentity {
  modelId?: string;
@@ -223,6 +229,7 @@ interface RawStateResponse {
    }
  >;
  runners?: Record<string, unknown>;
+  instanceLinks?: Record<string, RawInstanceLink>;
  downloads?: Record<string, unknown[]>;
  // New granular node state fields
  nodeIdentities?: Record<string, RawNodeIdentity>;
@@ -257,11 +264,12 @@ interface RawStateResponse {
 }

 export interface MessageAttachment {
-  type: "image" | "text" | "file" | "generated-image";
+  type: "image" | "text" | "file" | "generated-image" | "pdf";
  name: string;
  content?: string;
  preview?: string;
  mimeType?: string;
+  pageImages?: string[];
 }

 export interface TopLogprob {
@@ -540,6 +548,8 @@ class AppStore {
  topologyData = $state<TopologyData | null>(null);
  instances = $state<Record<string, unknown>>({});
  runners = $state<Record<string, unknown>>({});
+  instanceLinks = $state<Record<string, RawInstanceLink>>({});
+  featureFlags = $state<Record<string, boolean>>({});
  downloads = $state<Record<string, unknown[]>>({});
  nodeDisk = $state<
    Record<
@@ -1273,6 +1283,7 @@ class AppStore {

  startPolling() {
    this.fetchState();
+    this.fetchFeatureFlags();
    this.fetchInterval = setInterval(() => this.fetchState(), 1000);
  }

@@ -1284,6 +1295,16 @@ class AppStore {
    this.stopPreviewsPolling();
  }

+  async fetchFeatureFlags() {
+    try {
+      const response = await fetch("/v1/feature-flags");
+      if (!response.ok) return;
+      this.featureFlags = await response.json();
+    } catch {
+      // Silently ignore — defaults to all-disabled.
+    }
+  }
+
  async fetchState() {
    try {
      const response = await fetch("/state");
@@ -1309,6 +1330,11 @@ class AppStore {
      if (data.runners) {
        this.runners = data.runners;
      }
+      if (data.instanceLinks) {
+        this.instanceLinks = data.instanceLinks;
+      } else {
+        this.instanceLinks = {};
+      }
      if (data.downloads) {
        this.downloads = data.downloads;
      }
@@ -1669,7 +1695,15 @@ class AppStore {
              }
            }
          }
-          return { role: m.role, content: msgContent };
+          const out: {
+            role: string;
+            content: string;
+            reasoning_content?: string;
+          } = { role: m.role, content: msgContent };
+          if (m.role === "assistant" && m.thinking) {
+            out.reasoning_content = m.thinking;
+          }
+          return out;
        }),
      ];

@@ -1876,7 +1910,15 @@ class AppStore {
      const apiMessages = [
        systemPrompt,
        ...targetConversation.messages.slice(0, -1).map((m) => {
-          return { role: m.role, content: m.content };
+          const out: {
+            role: string;
+            content: string;
+            reasoning_content?: string;
+          } = { role: m.role, content: m.content };
+          if (m.role === "assistant" && m.thinking) {
+            out.reasoning_content = m.thinking;
+          }
+          return out;
        }),
      ];

@@ -2242,6 +2284,7 @@ class AppStore {
      type: string;
      textContent?: string;
      preview?: string;
+      pageImages?: string[];
    }[],
    enableThinking?: boolean | null,
  ): Promise<void> {
@@ -2277,6 +2320,20 @@ class AppStore {
            preview: file.preview,
            mimeType: file.type,
          });
+        } else if (
+          file.pageImages ||
+          (file.textContent && file.type === "application/pdf")
+        ) {
+          attachments.push({
+            type: "pdf",
+            name: file.name,
+            content: file.textContent,
+            pageImages: file.pageImages,
+            mimeType: file.type,
+          });
+          if (file.textContent) {
+            fileContext += `\n\n[File: ${file.name}]\n\`\`\`\n${file.textContent}\n\`\`\``;
+          }
        } else if (file.textContent) {
          attachments.push({
            type: "text",
@@ -2344,22 +2401,89 @@ class AppStore {
      const apiMessages = [
        systemPrompt,
        ...targetConversation.messages.slice(0, -1).map((m) => {
-          // Build content including any text file attachments
+          // Check if this message has image or PDF attachments
+          const visualAttachments = m.attachments?.filter(
+            (a) =>
+              (a.type === "image" && a.preview) ||
+              (a.type === "pdf" && a.pageImages?.length),
+          );
+
+          if (visualAttachments && visualAttachments.length > 0) {
+            // Build multimodal content array (OpenAI vision format)
+            const contentParts: Array<
+              | { type: "text"; text: string }
+              | { type: "image_url"; image_url: { url: string } }
+            > = [];
+
+            // Add image parts first
+            for (const att of visualAttachments) {
+              if (att.type === "image" && att.preview) {
+                contentParts.push({
+                  type: "image_url",
+                  image_url: { url: att.preview },
+                });
+              } else if (att.type === "pdf" && att.pageImages) {
+                for (const pageImg of att.pageImages) {
+                  contentParts.push({
+                    type: "image_url",
+                    image_url: { url: pageImg },
+                  });
+                }
+              }
+            }
+
+            // Build text content including any text/pdf file attachments
+            let textContent = m.content;
+            if (m.attachments) {
+              for (const attachment of m.attachments) {
+                if (
+                  (attachment.type === "text" || attachment.type === "pdf") &&
+                  attachment.content
+                ) {
+                  textContent += `\n\n[File: ${attachment.name}]\n\`\`\`\n${attachment.content}\n\`\`\``;
+                }
+              }
+            }
+
+            if (textContent) {
+              contentParts.push({ type: "text", text: textContent });
+            }
+
+            const out: {
+              role: string;
+              content: typeof contentParts;
+              reasoning_content?: string;
+            } = { role: m.role, content: contentParts };
+            if (m.role === "assistant" && m.thinking) {
+              out.reasoning_content = m.thinking;
+            }
+            return out;
+          }
+
+          // Text-only message (original path)
          let msgContent = m.content;

-          // Add text attachments as context
+          // Add text/pdf attachments as context
          if (m.attachments) {
            for (const attachment of m.attachments) {
-              if (attachment.type === "text" && attachment.content) {
+              if (
+                (attachment.type === "text" || attachment.type === "pdf") &&
+                attachment.content
+              ) {
                msgContent += `\n\n[File: ${attachment.name}]\n\`\`\`\n${attachment.content}\n\`\`\``;
              }
            }
          }

-          return {
-            role: m.role,
-            content: msgContent,
-          };
+          const out: {
+            role: string;
+            content: string;
+            reasoning_content?: string;
+          } = { role: m.role, content: msgContent };
+          if (m.role === "assistant" && m.thinking) {
+            out.reasoning_content = m.thinking;
+          }
+          return out;
        }),
      ];

@@ -2650,6 +2774,9 @@ class AppStore {
    this.syncActiveMessagesIfNeeded(targetConversationId);
    this.saveConversationsToStorage();

+    const abortController = new AbortController();
+    this.currentAbortController = abortController;
+
    try {
      // Determine the model to use
      const model = this.getModelForRequest(modelId);
@@ -2704,6 +2831,7 @@ class AppStore {
          "Content-Type": "application/json",
        },
        body: JSON.stringify(requestBody),
+        signal: abortController.signal,
      });

      if (!response.ok) {
@@ -2843,14 +2971,27 @@ class AppStore {
        );
      }
    } catch (error) {
-      console.error("Error generating image:", error);
-      this.handleStreamingError(
-        error,
-        targetConversationId,
-        assistantMessage.id,
-        "Failed to generate image",
-      );
+      if (abortController.signal.aborted) {
+        this.updateConversationMessage(
+          targetConversationId,
+          assistantMessage.id,
+          (msg) => {
+            msg.content = "Cancelled";
+            msg.attachments = [];
+          },
+        );
+        this.syncActiveMessagesIfNeeded(targetConversationId);
+      } else {
+        console.error("Error generating image:", error);
+        this.handleStreamingError(
+          error,
+          targetConversationId,
+          assistantMessage.id,
+          "Failed to generate image",
+        );
+      }
    } finally {
+      this.currentAbortController = null;
      this.isLoading = false;
      this.saveConversationsToStorage();
    }
@@ -2914,6 +3055,9 @@ class AppStore {
    // Clear editing state
    this.editingImage = null;

+    const abortController = new AbortController();
+    this.currentAbortController = abortController;
+
    try {
      // Determine the model to use
      const model = this.getModelForRequest(modelId);
@@ -2975,6 +3119,7 @@ class AppStore {
      const apiResponse = await fetch("/v1/images/edits", {
        method: "POST",
        body: formData,
+        signal: abortController.signal,
      });

      if (!apiResponse.ok) {
@@ -3075,14 +3220,27 @@ class AppStore {
        );
      }
    } catch (error) {
-      console.error("Error editing image:", error);
-      this.handleStreamingError(
-        error,
-        targetConversationId,
-        assistantMessage.id,
-        "Failed to edit image",
-      );
+      if (abortController.signal.aborted) {
+        this.updateConversationMessage(
+          targetConversationId,
+          assistantMessage.id,
+          (msg) => {
+            msg.content = "Cancelled";
+            msg.attachments = [];
+          },
+        );
+        this.syncActiveMessagesIfNeeded(targetConversationId);
+      } else {
+        console.error("Error editing image:", error);
+        this.handleStreamingError(
+          error,
+          targetConversationId,
+          assistantMessage.id,
+          "Failed to edit image",
+        );
+      }
    } finally {
+      this.currentAbortController = null;
      this.isLoading = false;
      this.saveConversationsToStorage();
    }
@@ -3149,6 +3307,85 @@ class AppStore {
    }
  }

+  /**
+   * Cancel/pause an active download on a specific node
+   */
+  async cancelDownload(nodeId: string, modelId: string): Promise<void> {
+    try {
+      const response = await fetch("/download/cancel", {
+        method: "POST",
+        headers: { "Content-Type": "application/json" },
+        body: JSON.stringify({
+          targetNodeId: nodeId,
+          modelId: modelId,
+        }),
+      });
+      if (!response.ok) {
+        const errorText = await response.text();
+        throw new Error(
+          `Failed to cancel download: ${response.status} - ${errorText}`,
+        );
+      }
+    } catch (error) {
+      console.error("Error cancelling download:", error);
+      throw error;
+    }
+  }
+
+  async createInstanceLink(
+    prefillInstances: string[],
+    decodeInstances: string[],
+  ): Promise<void> {
+    const response = await fetch("/v1/instance-links", {
+      method: "POST",
+      headers: { "Content-Type": "application/json" },
+      body: JSON.stringify({
+        prefill_instances: prefillInstances,
+        decode_instances: decodeInstances,
+      }),
+    });
+    if (!response.ok) {
+      throw new Error(
+        `Failed to create instance link: ${response.status} ${await response.text()}`,
+      );
+    }
+  }
+
+  async updateInstanceLink(
+    linkId: string,
+    prefillInstances: string[],
+    decodeInstances: string[],
+  ): Promise<void> {
+    const response = await fetch(
+      `/v1/instance-links/${encodeURIComponent(linkId)}`,
+      {
+        method: "PUT",
+        headers: { "Content-Type": "application/json" },
+        body: JSON.stringify({
+          prefill_instances: prefillInstances,
+          decode_instances: decodeInstances,
+        }),
+      },
+    );
+    if (!response.ok) {
+      throw new Error(
+        `Failed to update instance link: ${response.status} ${await response.text()}`,
+      );
+    }
+  }
+
+  async deleteInstanceLink(linkId: string): Promise<void> {
+    const response = await fetch(
+      `/v1/instance-links/${encodeURIComponent(linkId)}`,
+      { method: "DELETE" },
+    );
+    if (!response.ok) {
+      throw new Error(
+        `Failed to delete instance link: ${response.status} ${await response.text()}`,
+      );
+    }
+  }
+
  /**
   * Delete a downloaded model from a specific node
   */
@@ -3247,6 +3484,19 @@ export const prefillProgress = () => appStore.prefillProgress;
 export const topologyData = () => appStore.topologyData;
 export const instances = () => appStore.instances;
 export const runners = () => appStore.runners;
+export const instanceLinks = () => appStore.instanceLinks;
+export const featureFlags = () => appStore.featureFlags;
+export const createInstanceLink = (
+  prefillInstances: string[],
+  decodeInstances: string[],
+) => appStore.createInstanceLink(prefillInstances, decodeInstances);
+export const updateInstanceLink = (
+  linkId: string,
+  prefillInstances: string[],
+  decodeInstances: string[],
+) => appStore.updateInstanceLink(linkId, prefillInstances, decodeInstances);
+export const deleteInstanceLink = (linkId: string) =>
+  appStore.deleteInstanceLink(linkId);
 export const downloads = () => appStore.downloads;
 export const nodeDisk = () => appStore.nodeDisk;
 export const placementPreviews = () => appStore.placementPreviews;
@@ -3271,6 +3521,7 @@ export const sendMessage = (
    type: string;
    textContent?: string;
    preview?: string;
+    pageImages?: string[];
  }[],
  enableThinking?: boolean | null,
 ) => appStore.sendMessage(content, files, enableThinking);
@@ -3369,6 +3620,8 @@ export const resetImageGenerationParams = () =>
 // Download actions
 export const startDownload = (nodeId: string, shardMetadata: object) =>
  appStore.startDownload(nodeId, shardMetadata);
+export const cancelDownload = (nodeId: string, modelId: string) =>
+  appStore.cancelDownload(nodeId, modelId);
 export const deleteDownload = (nodeId: string, modelId: string) =>
  appStore.deleteDownload(nodeId, modelId);

@@ -2,6 +2,41 @@
 * File attachment types for the chat interface
 */

+import { getDocument, GlobalWorkerOptions, version } from "pdfjs-dist";
+import type { DocumentInitParameters } from "pdfjs-dist/types/src/display/api";
+
+// Safari (through at least 18/26) does not implement
+// ReadableStream.prototype[Symbol.asyncIterator], which pdfjs-dist uses
+// internally in getTextContent(). Without this polyfill, `for await (const n
+// of stream)` throws "undefined is not a function" and PDF processing fails.
+if (
+  typeof ReadableStream !== "undefined" &&
+  !(ReadableStream.prototype as unknown as Record<symbol, unknown>)[
+    Symbol.asyncIterator
+  ]
+) {
+  (ReadableStream.prototype as unknown as Record<symbol, unknown>)[
+    Symbol.asyncIterator
+  ] = async function* (this: ReadableStream<unknown>) {
+    const reader = this.getReader();
+    try {
+      while (true) {
+        const { done, value } = await reader.read();
+        if (done) return;
+        yield value;
+      }
+    } finally {
+      reader.releaseLock();
+    }
+  };
+}
+
+GlobalWorkerOptions.workerSrc = `https://cdn.jsdelivr.net/npm/pdfjs-dist@${version}/build/pdf.worker.mjs`;
+
+const PDF_PAGE_SCALE = 2.0;
+const PDF_MAX_PAGES = 20;
+const PDF_MAX_TEXT_CHARS = 100_000;
+
 export interface ChatUploadedFile {
  id: string;
  name: string;
@@ -10,6 +45,7 @@ export interface ChatUploadedFile {
  file: File;
  preview?: string;
  textContent?: string;
+  pageImages?: string[];
 }

 export interface ChatAttachment {
@@ -194,6 +230,58 @@ export function readFileAsText(file: File): Promise<string> {
  });
 }

+async function extractPdfContent(
+  file: File,
+): Promise<{ text: string; pageImages: string[] }> {
+  const arrayBuffer = await file.arrayBuffer();
+  const pdf = await getDocument({
+    data: new Uint8Array(arrayBuffer),
+    useSystemFonts: true,
+  } as DocumentInitParameters).promise;
+
+  const numPages = Math.min(pdf.numPages, PDF_MAX_PAGES);
+  const pageTexts: string[] = [];
+  const pageImages: string[] = [];
+
+  for (let i = 1; i <= numPages; i++) {
+    const page = await pdf.getPage(i);
+
+    const content = await page.getTextContent();
+    const strings = content.items
+      .filter((item: any) => "str" in item)
+      .map((item: any) => item.str as string);
+    pageTexts.push(strings.join(" "));
+
+    const viewport = page.getViewport({ scale: PDF_PAGE_SCALE });
+    const canvas = new OffscreenCanvas(viewport.width, viewport.height);
+    const ctx = canvas.getContext("2d");
+    if (ctx) {
+      await page.render({ canvasContext: ctx as any, viewport }).promise;
+      const blob = await canvas.convertToBlob({
+        type: "image/jpeg",
+        quality: 0.8,
+      });
+      const reader = new FileReader();
+      const dataUrl = await new Promise<string>((resolve, reject) => {
+        reader.onload = () => resolve(reader.result as string);
+        reader.onerror = () => reject(reader.error);
+        reader.readAsDataURL(blob);
+      });
+      pageImages.push(dataUrl);
+    }
+  }
+
+  let text = pageTexts.join("\n\n").trim();
+  if (text.length > PDF_MAX_TEXT_CHARS) {
+    text = text.slice(0, PDF_MAX_TEXT_CHARS) + "\n\n[truncated]";
+  }
+  if (pdf.numPages > PDF_MAX_PAGES) {
+    text += `\n\n[showing ${PDF_MAX_PAGES} of ${pdf.numPages} pages]`;
+  }
+
+  return { text, pageImages };
+}
+
 /**
 * Process uploaded files into ChatUploadedFile format
 */
@@ -223,7 +311,12 @@ export async function processUploadedFiles(
        const textContent = await readFileAsText(file);
        results.push({ ...base, textContent });
      } else if (category === "pdf") {
-        results.push(base);
+        const { text, pageImages } = await extractPdfContent(file);
+        results.push({
+          ...base,
+          textContent: text || undefined,
+          pageImages: pageImages.length > 0 ? pageImages : undefined,
+        });
      } else if (category === "audio") {
        const preview = await readFileAsDataURL(file);
        results.push({ ...base, preview });
@@ -0,0 +1,55 @@
+export async function copyText(text: string): Promise<boolean> {
+  if (
+    typeof window !== "undefined" &&
+    window.isSecureContext &&
+    navigator.clipboard?.writeText
+  ) {
+    try {
+      await navigator.clipboard.writeText(text);
+      return true;
+    } catch {
+      // fall through to execCommand fallback
+    }
+  }
+
+  if (typeof document === "undefined") {
+    return false;
+  }
+
+  const textarea = document.createElement("textarea");
+  textarea.value = text;
+  textarea.setAttribute("readonly", "");
+  textarea.style.position = "fixed";
+  textarea.style.top = "0";
+  textarea.style.left = "0";
+  textarea.style.width = "1px";
+  textarea.style.height = "1px";
+  textarea.style.padding = "0";
+  textarea.style.border = "none";
+  textarea.style.outline = "none";
+  textarea.style.boxShadow = "none";
+  textarea.style.background = "transparent";
+  textarea.style.opacity = "0";
+  document.body.appendChild(textarea);
+
+  const previousSelection = document.getSelection();
+  const previousRange =
+    previousSelection && previousSelection.rangeCount > 0
+      ? previousSelection.getRangeAt(0)
+      : null;
+
+  try {
+    textarea.focus();
+    textarea.select();
+    textarea.setSelectionRange(0, text.length);
+    return document.execCommand("copy");
+  } catch {
+    return false;
+  } finally {
+    document.body.removeChild(textarea);
+    if (previousRange && previousSelection) {
+      previousSelection.removeAllRanges();
+      previousSelection.addRange(previousRange);
+    }
+  }
+}
@@ -0,0 +1,44 @@
+// Mirrors src/exo/shared/models/model_cards.py:derive_base_model
+const QUANT_SUFFIXES = new RegExp(
+  "[-_ ](?:MLX|MXFP[0-9]+|NVFP[0-9]+|GPTQ|AWQ|GGUF|fp16|bf16|fp8|int[0-9]+|[0-9]+(?:\\.[0-9]+)?bit|Q[0-9]+(?:_[A-Z0-9]+)?|gs[0-9]+)" +
+    "(?:[-_ ](?:MLX|Q[0-9]+|Int[0-9]+|[A-Z0-9]+|gs[0-9]+))*$",
+  "i",
+);
+
+function normalize(s: string): string {
+  return s
+    .replaceAll("-", " ")
+    .replaceAll("_", " ")
+    .replaceAll("  ", " ")
+    .trim();
+}
+
+export function deriveBaseModel(modelId: string): string {
+  const short = modelId.includes("/")
+    ? (modelId.split("/").pop() ?? modelId)
+    : modelId;
+  const stripped = short.replace(QUANT_SUFFIXES, "");
+  return normalize(stripped);
+}
+
+export function baseModelsCompatible(a: string, b: string): boolean {
+  return deriveBaseModel(a).toLowerCase() === deriveBaseModel(b).toLowerCase();
+}
+
+// Mirrors src/exo/shared/models/model_cards.py:derive_family
+export function deriveFamily(modelId: string): string {
+  const short = modelId.includes("/")
+    ? (modelId.split("/").pop() ?? modelId)
+    : modelId;
+  const stripped = short
+    .replace(QUANT_SUFFIXES, "")
+    .toLowerCase()
+    .replaceAll("_", "-");
+  const parts = stripped.split(/[-.]/);
+  const familyParts: string[] = [];
+  for (const p of parts) {
+    if (/^\d+$/.test(p) || /^\d+[bm]?$/i.test(p)) break;
+    familyParts.push(p);
+  }
+  return familyParts.length > 0 ? familyParts.join("-") : stripped;
+}
@@ -265,6 +265,7 @@

  let mounted = $state(false);
  let localNodeId = $state<string | null>(null);
+  let pendingFirefoxQuery = $state<string | null>(null); // ?q= param deferred until state loads

  // ── Onboarding wizard state ──
  const ONBOARDING_COMPLETE_KEY = "exo-onboarding-complete";
@@ -1309,6 +1310,20 @@
      return;
    }

+    // Firefox AI sidebar integration: handle ?q= query parameter
+    // Firefox's built-in AI sidebar (about:config: browser.ml.chat.enabled) sends
+    // the user's prompt as ?q=<URL-encoded prompt> to the configured provider URL.
+    // See: https://support.mozilla.org/en-US/kb/ai-chatbot
+    const queryParam = params.get("q");
+    if (queryParam) {
+      // Clean up the URL to prevent re-submission on page refresh
+      window.history.replaceState({}, "", window.location.pathname);
+      // Defer the auto-send until cluster state is loaded (topologyData,
+      // instances, availableMemory) so that model auto-selection works
+      // correctly. The $effect below will pick this up once data is ready.
+      pendingFirefoxQuery = queryParam;
+    }
+
    // Check server-side onboarding state (persisted in ~/.exo)
    try {
      const res = await fetch("/onboarding");
@@ -1329,6 +1344,18 @@
    }
  });

+  // Deferred Firefox AI sidebar auto-send: wait for cluster state and model
+  // list before submitting. Both data (from /state polling) and models (from
+  // the async /models fetch in onMount) must be loaded for handleAutoSend to
+  // correctly auto-select a model.
+  $effect(() => {
+    if (pendingFirefoxQuery && data && models.length > 0) {
+      const query = pendingFirefoxQuery;
+      pendingFirefoxQuery = null;
+      handleChatSend(query);
+    }
+  });
+
  async function fetchModels() {
    try {
      const response = await fetch("/models");
@@ -3408,6 +3435,7 @@
            >
              <li>Connect nodes with TB5 cables</li>
              <li>Boot to Recovery (hold power 10s → Options)</li>
+              <li>Open Terminal from the Utilities menu</li>
              <li>
                Run
                <code class="text-yellow-300 bg-yellow-400/10 px-1 rounded"
@@ -4795,6 +4823,7 @@
                >
                  <li>Connect nodes with TB5 cables</li>
                  <li>Boot to Recovery (hold power 10s → Options)</li>
+                  <li>Open Terminal from the Utilities menu</li>
                  <li>
                    Run
                    <code class="text-yellow-300 bg-yellow-400/10 px-1 rounded"
@@ -4941,6 +4970,7 @@
                  >
                    <li>Connect nodes with TB5 cables</li>
                    <li>Boot to Recovery (hold power 10s → Options)</li>
+                    <li>Open Terminal from the Utilities menu</li>
                    <li>
                      Run
                      <code
@@ -0,0 +1,81 @@
+<script lang="ts">
+  import { browser } from "$app/environment";
+  import HeaderNav from "$lib/components/HeaderNav.svelte";
+  import PrefillDecodeDisaggregation from "$lib/components/PrefillDecodeDisaggregation.svelte";
+  import { featureFlags, refreshState } from "$lib/stores/app.svelte";
+  import { onMount } from "svelte";
+
+  type TabId = "prefill-decode";
+
+  const tabs: { id: TabId; label: string }[] = [
+    { id: "prefill-decode", label: "Prefill / Decode" },
+  ];
+
+  let activeTab = $state<TabId>(tabs[0].id);
+  let flagsLoaded = $state(false);
+
+  onMount(() => {
+    refreshState().finally(() => {
+      flagsLoaded = true;
+    });
+  });
+
+  const flags = $derived(featureFlags());
+  const enabled = $derived(flags["disaggregation"] === true);
+
+  $effect(() => {
+    if (browser && flagsLoaded && !enabled) {
+      // No advanced features enabled — bounce home.
+      window.location.hash = "/";
+    }
+  });
+</script>
+
+<div class="min-h-screen bg-exo-dark-gray flex flex-col">
+  <HeaderNav />
+
+  <main class="flex-1 max-w-[1100px] mx-auto w-full px-4 md:px-6 py-8">
+    {#if !flagsLoaded}
+      <div class="text-exo-light-gray/60 text-sm">Loading…</div>
+    {:else if !enabled}
+      <div class="text-exo-light-gray/60 text-sm">
+        No advanced features enabled. Set <code
+          class="text-exo-yellow font-mono">ENABLE_DISAGGREGATION=true</code
+        > on the cluster to access prefill/decode disaggregation.
+      </div>
+    {:else}
+      <div class="mb-4">
+        <h1
+          class="text-white text-xl md:text-2xl font-semibold tracking-wide mb-2"
+        >
+          Advanced
+        </h1>
+        <p class="text-exo-light-gray/60 text-sm">
+          Cluster-level configuration. Most users don't need anything here.
+        </p>
+      </div>
+
+      <div
+        class="flex flex-wrap gap-2 mb-6 border-b border-exo-light-gray/10 pb-3"
+      >
+        {#each tabs as tab (tab.id)}
+          <button
+            onclick={() => (activeTab = tab.id)}
+            class="px-3 py-1.5 text-xs rounded-md transition-all cursor-pointer
+              {activeTab === tab.id
+              ? 'bg-exo-yellow/15 text-exo-yellow border border-exo-yellow/30'
+              : 'text-exo-light-gray/60 hover:text-white/80 border border-transparent hover:border-exo-light-gray/20'}"
+          >
+            {tab.label}
+          </button>
+        {/each}
+      </div>
+
+      <div class="space-y-4">
+        {#if activeTab === "prefill-decode"}
+          <PrefillDecodeDisaggregation />
+        {/if}
+      </div>
+    {/if}
+  </main>
+</div>
@@ -9,6 +9,7 @@
    refreshState,
    lastUpdate as lastUpdateStore,
    startDownload,
+    cancelDownload,
    deleteDownload,
  } from "$lib/stores/app.svelte";
  import {
@@ -349,6 +350,59 @@
  });
 </script>

+{#snippet trashIcon()}
+  <svg
+    class="w-5 h-5"
+    viewBox="0 0 20 20"
+    fill="none"
+    stroke="currentColor"
+    stroke-width="2"
+  >
+    <path
+      d="M4 6h12M8 6V4h4v2m1 0v10a1 1 0 01-1 1H8a1 1 0 01-1-1V6h6"
+      stroke-linecap="round"
+      stroke-linejoin="round"
+    ></path>
+  </svg>
+{/snippet}
+
+{#snippet downloadIcon(size?: string)}
+  <svg
+    class={size ?? "w-5 h-5"}
+    viewBox="0 0 20 20"
+    fill="none"
+    stroke="currentColor"
+    stroke-width="2"
+  >
+    <path
+      d="M10 3v10m0 0l-3-3m3 3l3-3M3 17h14"
+      stroke-linecap="round"
+      stroke-linejoin="round"
+    ></path>
+  </svg>
+{/snippet}
+
+{#snippet pauseIcon()}
+  <svg class="w-5 h-5" viewBox="0 0 20 20" fill="currentColor">
+    <path
+      fill-rule="evenodd"
+      d="M6 4h2v12H6V4zm6 0h2v12h-2V4z"
+      clip-rule="evenodd"
+    ></path>
+  </svg>
+{/snippet}
+
+{#snippet deleteButton(nodeId: string, modelId: string)}
+  <button
+    type="button"
+    class="text-white/50 hover:text-red-400 transition-colors cursor-pointer"
+    onclick={() => deleteDownload(nodeId, modelId)}
+    title="Delete from this node"
+  >
+    {@render trashIcon()}
+  </button>
+{/snippet}
+
 <div class="min-h-screen bg-exo-dark-gray text-white">
  <HeaderNav showHome={true} />
  <div class="max-w-7xl mx-auto px-4 lg:px-8 py-6 space-y-6">
@@ -486,27 +540,7 @@
                        <span class="text-xs text-white/70"
                          >{formatBytes(cell.totalBytes)}</span
                        >
-                        <button
-                          type="button"
-                          class="text-white/50 hover:text-red-400 transition-colors mt-0.5 cursor-pointer"
-                          onclick={() =>
-                            deleteDownload(col.nodeId, row.modelId)}
-                          title="Delete from this node"
-                        >
-                          <svg
-                            class="w-5 h-5"
-                            viewBox="0 0 20 20"
-                            fill="none"
-                            stroke="currentColor"
-                            stroke-width="2"
-                          >
-                            <path
-                              d="M4 6h12M8 6V4h4v2m1 0v10a1 1 0 01-1 1H8a1 1 0 01-1-1V6h6"
-                              stroke-linecap="round"
-                              stroke-linejoin="round"
-                            ></path>
-                          </svg>
-                        </button>
+                        {@render deleteButton(col.nodeId, row.modelId)}
                      </div>
                    {:else if cell.kind === "downloading"}
                      <div
@@ -533,6 +567,18 @@
                        <span class="text-[10px] text-white/70"
                          >{formatSpeed(cell.speed)}</span
                        >
+                        <div class="flex gap-1 mt-0.5">
+                          <button
+                            type="button"
+                            class="text-white/50 hover:text-exo-yellow transition-colors cursor-pointer"
+                            onclick={() =>
+                              cancelDownload(col.nodeId, row.modelId)}
+                            title="Pause download"
+                          >
+                            {@render pauseIcon()}
+                          </button>
+                          {@render deleteButton(col.nodeId, row.modelId)}
+                        </div>
                      </div>
                    {:else if cell.kind === "pending"}
                      <div
@@ -558,32 +604,24 @@
                              ).toFixed(1)}%"
                            ></div>
                          </div>
-                          {#if row.shardMetadata}
-                            <button
-                              type="button"
-                              class="text-white/50 hover:text-exo-yellow transition-colors cursor-pointer"
-                              onclick={() =>
-                                startDownload(col.nodeId, row.shardMetadata!)}
-                              title="Resume download on this node"
-                            >
-                              <svg
-                                class="w-5 h-5"
-                                viewBox="0 0 20 20"
-                                fill="none"
-                                stroke="currentColor"
-                                stroke-width="2"
+                          <div class="flex gap-1">
+                            {#if row.shardMetadata}
+                              <button
+                                type="button"
+                                class="text-white/50 hover:text-exo-yellow transition-colors cursor-pointer"
+                                onclick={() =>
+                                  startDownload(col.nodeId, row.shardMetadata!)}
+                                title="Resume download on this node"
                              >
-                                <path
-                                  d="M10 3v10m0 0l-3-3m3 3l3-3M3 17h14"
-                                  stroke-linecap="round"
-                                  stroke-linejoin="round"
-                                ></path>
-                              </svg>
-                            </button>
-                          {:else}
-                            <span class="text-white/50 text-[10px]">paused</span
-                            >
-                          {/if}
+                                {@render downloadIcon()}
+                              </button>
+                            {:else}
+                              <span class="text-white/50 text-[10px]"
+                                >paused</span
+                              >
+                            {/if}
+                            {@render deleteButton(col.nodeId, row.modelId)}
+                          </div>
                        {:else if row.shardMetadata}
                          <button
                            type="button"
@@ -592,19 +630,7 @@
                              startDownload(col.nodeId, row.shardMetadata!)}
                            title="Start download on this node"
                          >
-                            <svg
-                              class="w-6 h-6"
-                              viewBox="0 0 20 20"
-                              fill="none"
-                              stroke="currentColor"
-                              stroke-width="2"
-                            >
-                              <path
-                                d="M10 3v10m0 0l-3-3m3 3l3-3M3 17h14"
-                                stroke-linecap="round"
-                                stroke-linejoin="round"
-                              ></path>
-                            </svg>
+                            {@render downloadIcon("w-6 h-6")}
                          </button>
                        {:else}
                          <span class="text-white/40 text-sm">...</span>
@@ -626,29 +652,20 @@
                            clip-rule="evenodd"
                          ></path>
                        </svg>
-                        {#if row.shardMetadata}
-                          <button
-                            type="button"
-                            class="text-white/50 hover:text-exo-yellow transition-colors cursor-pointer"
-                            onclick={() =>
-                              startDownload(col.nodeId, row.shardMetadata!)}
-                            title="Retry download on this node"
-                          >
-                            <svg
-                              class="w-5 h-5"
-                              viewBox="0 0 20 20"
-                              fill="none"
-                              stroke="currentColor"
-                              stroke-width="2"
+                        <div class="flex gap-1">
+                          {#if row.shardMetadata}
+                            <button
+                              type="button"
+                              class="text-white/50 hover:text-exo-yellow transition-colors cursor-pointer"
+                              onclick={() =>
+                                startDownload(col.nodeId, row.shardMetadata!)}
+                              title="Retry download on this node"
                            >
-                              <path
-                                d="M10 3v10m0 0l-3-3m3 3l3-3M3 17h14"
-                                stroke-linecap="round"
-                                stroke-linejoin="round"
-                              ></path>
-                            </svg>
-                          </button>
-                        {/if}
+                              {@render downloadIcon()}
+                            </button>
+                          {/if}
+                          {@render deleteButton(col.nodeId, row.modelId)}
+                        </div>
                      </div>
                    {:else}
                      <div
@@ -666,19 +683,7 @@
                              startDownload(col.nodeId, row.shardMetadata!)}
                            title="Download to this node"
                          >
-                            <svg
-                              class="w-5 h-5"
-                              viewBox="0 0 20 20"
-                              fill="none"
-                              stroke="currentColor"
-                              stroke-width="2"
-                            >
-                              <path
-                                d="M10 3v10m0 0l-3-3m3 3l3-3M3 17h14"
-                                stroke-linecap="round"
-                                stroke-linejoin="round"
-                              ></path>
-                            </svg>
+                            {@render downloadIcon()}
                          </button>
                        {/if}
                      </div>
@@ -0,0 +1,688 @@
+<script lang="ts">
+  import { browser } from "$app/environment";
+  import { fade } from "svelte/transition";
+  import HeaderNav from "$lib/components/HeaderNav.svelte";
+  import IntegrationCard from "$lib/components/IntegrationCard.svelte";
+  import { instances, refreshState } from "$lib/stores/app.svelte";
+  import { onMount } from "svelte";
+
+  const apiUrl = browser
+    ? window.location.origin.replace("localhost", "127.0.0.1")
+    : "http://127.0.0.1:52415";
+
+  const instancesData = $derived(instances());
+
+  let modelCapabilities = $state<Record<string, string[]>>({});
+  let modelContextLengths = $state<Record<string, number>>({});
+  let modelReasoningDialects = $state<Record<string, string>>({});
+
+  const runningModels = $derived.by(() => {
+    const models: string[] = [];
+    for (const [, wrapper] of Object.entries(instancesData)) {
+      if (wrapper && typeof wrapper === "object") {
+        const values = Object.values(wrapper as Record<string, unknown>);
+        if (values.length > 0) {
+          const instance = values[0];
+          if (instance && typeof instance === "object") {
+            const inst = instance as {
+              shardAssignments?: { modelId?: string };
+            };
+            const modelId = inst.shardAssignments?.modelId;
+            if (modelId && !models.includes(modelId)) {
+              models.push(modelId);
+            }
+          }
+        }
+      }
+    }
+    return models;
+  });
+
+  function estimateParamSize(modelId: string): number {
+    const match = modelId.match(/(\d+(?:\.\d+)?)[Bb]/);
+    return match ? parseFloat(match[1]) : 0;
+  }
+
+  const modelsBySize = $derived(
+    [...runningModels].sort(
+      (a, b) => estimateParamSize(b) - estimateParamSize(a),
+    ),
+  );
+
+  const defaultTiers = $derived.by(() => {
+    const n = modelsBySize.length;
+    if (n === 0)
+      return {
+        opus: "your-model-id",
+        sonnet: "your-model-id",
+        haiku: "your-model-id",
+      };
+    if (n === 1)
+      return {
+        opus: modelsBySize[0],
+        sonnet: modelsBySize[0],
+        haiku: modelsBySize[0],
+      };
+    if (n === 2)
+      return {
+        opus: modelsBySize[0],
+        sonnet: modelsBySize[1],
+        haiku: modelsBySize[1],
+      };
+    return {
+      opus: modelsBySize[0],
+      sonnet: modelsBySize[Math.floor(n / 2)],
+      haiku: modelsBySize[n - 1],
+    };
+  });
+
+  let opusModel = $state("");
+  let sonnetModel = $state("");
+  let haikuModel = $state("");
+
+  $effect(() => {
+    opusModel = defaultTiers.opus;
+    sonnetModel = defaultTiers.sonnet;
+    haikuModel = defaultTiers.haiku;
+  });
+
+  let codexModel = $state("");
+  let codexMcpPath = $state("/Users/username");
+  let openClawModel = $state("");
+  let piModel = $state("");
+  $effect(() => {
+    const def = modelsBySize.length > 0 ? modelsBySize[0] : "your-model-id";
+    codexModel = def;
+    openClawModel = def;
+    piModel = def;
+  });
+
+  const claudeShellCommand = $derived(
+    [
+      `ANTHROPIC_BASE_URL=${apiUrl} \\`,
+      `ANTHROPIC_API_KEY=x \\`,
+      `ANTHROPIC_DEFAULT_OPUS_MODEL=${opusModel} \\`,
+      `ANTHROPIC_DEFAULT_SONNET_MODEL=${sonnetModel} \\`,
+      `ANTHROPIC_DEFAULT_HAIKU_MODEL=${haikuModel} \\`,
+      `API_TIMEOUT_MS=3000000 \\`,
+      `CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 \\`,
+      `claude`,
+    ].join("\n"),
+  );
+
+  const claudeSettingsJson = $derived(
+    JSON.stringify(
+      {
+        env: {
+          ANTHROPIC_BASE_URL: apiUrl,
+          ANTHROPIC_API_KEY: "x",
+          ANTHROPIC_DEFAULT_OPUS_MODEL: opusModel,
+          ANTHROPIC_DEFAULT_SONNET_MODEL: sonnetModel,
+          ANTHROPIC_DEFAULT_HAIKU_MODEL: haikuModel,
+          API_TIMEOUT_MS: "3000000",
+          CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC: "1",
+        },
+      },
+      null,
+      2,
+    ),
+  );
+
+  const openCodeConfig = $derived.by(() => {
+    const models: Record<string, Record<string, unknown>> = {};
+    for (const modelId of runningModels) {
+      const caps = modelCapabilities[modelId] || [];
+      const ctxLen = modelContextLengths[modelId] || 0;
+      const dialect = modelReasoningDialects[modelId];
+      const entry: Record<string, unknown> = { name: modelId };
+      if (ctxLen > 0) {
+        entry.limit = { context: ctxLen, output: Math.min(ctxLen, 16384) };
+      }
+      if (caps.includes("vision")) {
+        entry.modalities = { input: ["text", "image"], output: ["text"] };
+      }
+      // Reasoning round-trip: opencode's `interleaved` field tells the
+      // openai-compatible adapter to send the assistant's prior
+      // reasoning_content back in subsequent turns. Emit it for dialects
+      // whose chat templates use prior reasoning:
+      //   - `tool_conditional` (DeepSeek V3.2 / V4): wrapper preserves all
+      //     reasoning when tools are present.
+      //   - `post_last_user` (Qwen3-Thinking, GLM 4.5+, MiniMax M2.x):
+      //     Jinja template reads reasoning_content for assistant turns since
+      //     the last user message — exactly the tool-chain window.
+      //   - `channel` (gpt-oss / Harmony): the model's Jinja template reads
+      //     `message.thinking` rather than `message.reasoning_content`, but
+      //     the server bridges `reasoning_content` → `thinking` before
+      //     rendering, so the round-trip works through the standard field.
+      // `suffix` (Kimi): reasoning lives in content; no separate field path.
+      if (
+        dialect === "tool_conditional" ||
+        dialect === "post_last_user" ||
+        dialect === "channel"
+      ) {
+        entry.interleaved = { field: "reasoning_content" };
+      }
+      models[modelId] = entry;
+    }
+    if (Object.keys(models).length === 0) {
+      models["your-model-id"] = { name: "your-model-name" };
+    }
+    const firstModel =
+      runningModels.length > 0 ? runningModels[0] : "your-model-id";
+    return JSON.stringify(
+      {
+        $schema: "https://opencode.ai/config.json",
+        provider: {
+          exo: {
+            npm: "@ai-sdk/openai-compatible",
+            name: "exo",
+            options: {
+              baseURL: `${apiUrl}/v1`,
+              apiKey: "x",
+            },
+            models,
+          },
+        },
+        model: `exo/${firstModel}`,
+      },
+      null,
+      2,
+    );
+  });
+
+  const codexShellCommand = $derived(`EXO_API_KEY=x npx @openai/codex`);
+
+  const codexConfig = $derived(
+    [
+      `model = "${codexModel}"`,
+      `model_provider = "exo"`,
+      ``,
+      `[model_providers.exo]`,
+      `name = "exo"`,
+      `base_url = "${apiUrl}/v1"`,
+      `env_key = "EXO_API_KEY"`,
+      ``,
+      `[mcp_servers.filesystem]`,
+      `command = "npx"`,
+      `args = ["-y", "@modelcontextprotocol/server-filesystem", "${codexMcpPath}"]`,
+    ].join("\n"),
+  );
+
+  const openClawConfig = $derived(
+    JSON.stringify(
+      {
+        gateway: { mode: "local" },
+        models: {
+          providers: {
+            exo: {
+              baseUrl: `${apiUrl}/v1`,
+              apiKey: "x",
+              api: "openai-completions",
+              models: [
+                {
+                  id: openClawModel,
+                  name: "exo local",
+                  input: (modelCapabilities[openClawModel] || []).includes(
+                    "vision",
+                  )
+                    ? ["text", "image"]
+                    : ["text"],
+                },
+              ],
+            },
+          },
+        },
+        agents: {
+          defaults: {
+            model: `exo/${openClawModel}`,
+          },
+        },
+      },
+      null,
+      2,
+    ),
+  );
+
+  const piModelsJson = $derived.by(() => {
+    const models: Record<string, unknown>[] = [];
+    for (const modelId of runningModels) {
+      const caps = modelCapabilities[modelId] || [];
+      const ctxLen = modelContextLengths[modelId] || 0;
+      const entry: Record<string, unknown> = { id: modelId };
+      if (caps.includes("vision")) {
+        entry.input = ["text", "image"];
+      }
+      // Mark thinking-capable models so pi surfaces its thinking-level selector
+      // for them. exo capability strings: "thinking" (model emits reasoning
+      // content) and "thinking_toggle" (user can turn it on/off).
+      if (caps.includes("thinking") || caps.includes("thinking_toggle")) {
+        entry.reasoning = true;
+      }
+      if (ctxLen > 0) {
+        entry.contextWindow = ctxLen;
+      }
+      models.push(entry);
+    }
+    if (models.length === 0) {
+      models.push({ id: "your-model-id" });
+    }
+    return JSON.stringify(
+      {
+        providers: {
+          exo: {
+            baseUrl: `${apiUrl}/v1`,
+            api: "openai-completions",
+            apiKey: "exo",
+            compat: {
+              supportsDeveloperRole: false,
+              // exo's OpenAI surface takes a boolean `enable_thinking` toggle,
+              // not graded effort levels, so disable pi's `reasoning_effort`
+              // parameter and use the matching top-level-boolean format.
+              supportsReasoningEffort: false,
+              thinkingFormat: "qwen",
+            },
+            models,
+          },
+        },
+      },
+      null,
+      2,
+    );
+  });
+
+  const piShellCommand = $derived(`pi --provider exo --model ${piModel}`);
+
+  const ollamaCommand = $derived(
+    `OLLAMA_HOST=${apiUrl}/ollama ollama run ${modelsBySize.length > 0 ? modelsBySize[0] : "your-model-id"}`,
+  );
+
+  const openWebUiCommand = $derived(
+    [
+      `docker run -d -p 3000:8080 \\`,
+      `  -e OLLAMA_BASE_URL=${apiUrl.replace("localhost", "host.docker.internal")}/ollama \\`,
+      `  -v open-webui:/app/backend/data \\`,
+      `  --name open-webui \\`,
+      `  ghcr.io/open-webui/open-webui:main`,
+    ].join("\n"),
+  );
+
+  const n8nDockerCommand = $derived(
+    [
+      `docker run -d -p 5678:5678 \\`,
+      `  -v n8n_data:/home/node/.n8n \\`,
+      `  --name n8n \\`,
+      `  docker.n8n.io/n8nio/n8n`,
+    ].join("\n"),
+  );
+
+  const n8nCredentialSteps = $derived(
+    [
+      `1. Go to Credentials → Add Credential → search "OpenAI API"`,
+      `2. Set API Key to: x`,
+      `3. Set Base URL to: ${apiUrl.replace("127.0.0.1", "host.docker.internal").replace("localhost", "host.docker.internal")}/v1`,
+      `4. Save the credential`,
+    ].join("\n"),
+  );
+
+  const n8nWorkflowSteps = $derived(
+    [
+      `1. Create a new workflow → "Start from Scratch"`,
+      `2. Add an "AI Agent" or "Basic LLM Chain" node`,
+      `3. Inside it, add an "OpenAI Chat Model" sub-node`,
+      `4. Select the OpenAI credential you just created`,
+      `5. Set Model to "From list" and pick your model (e.g. ${modelsBySize.length > 0 ? modelsBySize[0] : "your-model-id"})`,
+      `6. Optionally toggle "Use Responses API", add Built-in Tools, or click "Add Option" for sampling settings`,
+      `7. Connect a "Chat Trigger" node for interactive chat`,
+      `8. On the Chat Trigger, enable "Allow File Uploads" for vision`,
+    ].join("\n"),
+  );
+
+  const firefoxConfig = $derived(
+    [
+      `1. Open about:config in Firefox`,
+      `2. Set browser.ml.chat.enabled to true`,
+      `3. Set browser.ml.chat.hideLocalhost to false`,
+      `4. Set browser.ml.chat.provider to: ${apiUrl}/`,
+    ].join("\n"),
+  );
+
+  const tabs = [
+    "Claude Code",
+    "OpenCode",
+    "Codex",
+    "OpenClaw",
+    "Pi",
+    "Open WebUI",
+    "n8n",
+    "Firefox",
+  ] as const;
+  type Tab = (typeof tabs)[number];
+  const stored = browser ? localStorage.getItem("exo-integrations-tab") : null;
+  let activeTab = $state<Tab>(
+    stored && tabs.includes(stored as Tab) ? (stored as Tab) : "Claude Code",
+  );
+  $effect(() => {
+    if (browser) localStorage.setItem("exo-integrations-tab", activeTab);
+  });
+
+  const selectClass =
+    "bg-black/30 border border-exo-light-gray/20 rounded px-2 py-1.5 text-white font-mono text-xs focus:border-exo-yellow/50 focus:outline-none appearance-none cursor-pointer";
+
+  onMount(async () => {
+    refreshState();
+    try {
+      const resp = await fetch("/v1/models");
+      const data = (await resp.json()) as {
+        data: {
+          id: string;
+          capabilities: string[];
+          context_length: number;
+          reasoning_dialect?: string;
+        }[];
+      };
+      const caps: Record<string, string[]> = {};
+      const ctxs: Record<string, number> = {};
+      const dialects: Record<string, string> = {};
+      for (const model of data.data) {
+        caps[model.id] = model.capabilities || [];
+        if (model.context_length > 0) ctxs[model.id] = model.context_length;
+        if (model.reasoning_dialect)
+          dialects[model.id] = model.reasoning_dialect;
+      }
+      modelCapabilities = caps;
+      modelContextLengths = ctxs;
+      modelReasoningDialects = dialects;
+    } catch {
+      /* ignore */
+    }
+  });
+</script>
+
+<div class="min-h-screen bg-exo-dark-gray flex flex-col">
+  <HeaderNav showHome={true} />
+
+  <main
+    class="flex-1 max-w-3xl mx-auto w-full px-4 md:px-6 py-8"
+    in:fade={{ duration: 200 }}
+  >
+    <div class="mb-8">
+      <h1
+        class="text-white text-xl md:text-2xl font-semibold tracking-wide mb-2"
+      >
+        Integrations
+      </h1>
+      <p class="text-exo-light-gray/60 text-sm">
+        Connect external tools to your exo cluster.
+      </p>
+    </div>
+
+    <!-- Status -->
+    <div class="mb-8">
+      <span class="text-exo-light-gray/70 text-xs uppercase tracking-wider"
+        >API Endpoint</span
+      >
+      <span class="text-white font-mono text-sm ml-2">{apiUrl}</span>
+      {#if runningModels.length > 0}
+        <div class="text-exo-light-gray/50 text-xs mt-2">
+          Running model{runningModels.length > 1 ? "s" : ""}:
+          <ul class="mt-1 space-y-0.5 list-none">
+            {#each runningModels as model}
+              <li class="text-exo-yellow font-mono">{model}</li>
+            {/each}
+          </ul>
+        </div>
+      {:else}
+        <p class="text-exo-light-gray/40 text-xs mt-2 italic">
+          No models currently running
+        </p>
+      {/if}
+    </div>
+
+    <!-- API Endpoints -->
+    <div class="mb-8">
+      <div
+        class="flex flex-col sm:flex-row gap-3 text-xs font-mono text-exo-light-gray/70"
+      >
+        <div
+          class="flex-1 bg-black/20 border border-exo-light-gray/10 rounded px-3 py-2"
+        >
+          <span class="text-exo-light-gray/40 text-[10px] uppercase block mb-1"
+            >OpenAI-compatible</span
+          >
+          <span class="text-white/80">{apiUrl}/v1</span>
+        </div>
+        <div
+          class="flex-1 bg-black/20 border border-exo-light-gray/10 rounded px-3 py-2"
+        >
+          <span class="text-exo-light-gray/40 text-[10px] uppercase block mb-1"
+            >Claude-compatible</span
+          >
+          <span class="text-white/80">{apiUrl}</span>
+        </div>
+        <div
+          class="flex-1 bg-black/20 border border-exo-light-gray/10 rounded px-3 py-2"
+        >
+          <span class="text-exo-light-gray/40 text-[10px] uppercase block mb-1"
+            >Ollama-compatible</span
+          >
+          <span class="text-white/80">{apiUrl}/ollama</span>
+        </div>
+      </div>
+    </div>
+
+    <!-- Tabs -->
+    <div
+      class="flex flex-wrap gap-2 mb-6 border-b border-exo-light-gray/10 pb-3"
+    >
+      {#each tabs as tab}
+        <button
+          onclick={() => (activeTab = tab)}
+          class="px-3 py-1.5 text-xs rounded-md transition-all cursor-pointer
+            {activeTab === tab
+            ? 'bg-exo-yellow/15 text-exo-yellow border border-exo-yellow/30'
+            : 'text-exo-light-gray/60 hover:text-white/80 border border-transparent hover:border-exo-light-gray/20'}"
+        >
+          {tab}
+        </button>
+      {/each}
+    </div>
+
+    <!-- Tab Content -->
+    <div class="space-y-4">
+      {#if activeTab === "Claude Code"}
+        {#if runningModels.length > 1}
+          <div class="grid grid-cols-3 gap-3 text-xs">
+            {#each [{ label: "Opus", bind: () => opusModel, set: (v: string) => (opusModel = v) }, { label: "Sonnet", bind: () => sonnetModel, set: (v: string) => (sonnetModel = v) }, { label: "Haiku", bind: () => haikuModel, set: (v: string) => (haikuModel = v) }] as tier}
+              <div>
+                <span
+                  class="text-exo-light-gray/50 text-[10px] uppercase tracking-wider block mb-1"
+                  >{tier.label}</span
+                >
+                <select
+                  value={tier.bind()}
+                  onchange={(e) =>
+                    tier.set((e.target as HTMLSelectElement).value)}
+                  class="w-full {selectClass}"
+                >
+                  {#each runningModels as model}
+                    <option value={model}>{model.split("/").pop()}</option>
+                  {/each}
+                </select>
+              </div>
+            {/each}
+          </div>
+        {/if}
+        <IntegrationCard
+          title="Shell Command"
+          subtitle="Run in terminal"
+          description="Launch Claude Code with exo as the backend. Paste this into your terminal."
+          config={claudeShellCommand}
+          language="bash"
+        />
+        <IntegrationCard
+          title="Settings File"
+          subtitle="~/.claude/settings.json"
+          description="Or add this to your Claude Code settings for persistent configuration."
+          config={claudeSettingsJson}
+        />
+      {:else if activeTab === "OpenCode"}
+        <IntegrationCard
+          title="Config File"
+          subtitle="opencode.json"
+          description="Add this to your project root or ~/.config/opencode/opencode.json for global config. Vision models automatically get image input modality."
+          config={openCodeConfig}
+        />
+      {:else if activeTab === "Codex"}
+        <div class="flex gap-3 text-xs">
+          {#if runningModels.length > 1}
+            <div>
+              <span
+                class="text-exo-light-gray/50 text-[10px] uppercase tracking-wider block mb-1"
+                >Model</span
+              >
+              <select bind:value={codexModel} class={selectClass}>
+                {#each runningModels as model}
+                  <option value={model}>{model.split("/").pop()}</option>
+                {/each}
+              </select>
+            </div>
+          {/if}
+          <div class="flex-1">
+            <span
+              class="text-exo-light-gray/50 text-[10px] uppercase tracking-wider block mb-1"
+              >MCP Filesystem Path</span
+            >
+            <input
+              type="text"
+              bind:value={codexMcpPath}
+              class="w-full bg-black/30 border border-exo-light-gray/20 rounded px-2 py-1.5 text-white font-mono text-xs focus:border-exo-yellow/50 focus:outline-none"
+            />
+          </div>
+        </div>
+        <IntegrationCard
+          title="Config File"
+          subtitle="~/.codex/config.toml"
+          description="Add this to your Codex CLI config so the model and provider persist."
+          config={codexConfig}
+        />
+        <IntegrationCard
+          title="Shell Command"
+          subtitle="Run in terminal"
+          description="Launch Codex with exo as the backend."
+          config={codexShellCommand}
+          language="bash"
+        />
+      {:else if activeTab === "OpenClaw"}
+        {#if runningModels.length > 1}
+          <div class="text-xs">
+            <span
+              class="text-exo-light-gray/50 text-[10px] uppercase tracking-wider block mb-1"
+              >Model</span
+            >
+            <select bind:value={openClawModel} class={selectClass}>
+              {#each runningModels as model}
+                <option value={model}>{model.split("/").pop()}</option>
+              {/each}
+            </select>
+          </div>
+        {/if}
+        <IntegrationCard
+          title="Config File"
+          subtitle="~/.openclaw/openclaw.json"
+          description="Add this to your OpenClaw config. If you haven't installed OpenClaw yet, run: npm install -g openclaw@latest"
+          config={openClawConfig}
+        />
+        <IntegrationCard
+          title="Setup Commands"
+          subtitle="Run in terminal"
+          description="After saving the config, run these commands to fix metadata and start the gateway."
+          config={`openclaw doctor --fix${(modelCapabilities[openClawModel] || []).includes("vision") ? `\nopenclaw models set-image exo/${openClawModel}` : ""}\nopenclaw gateway &\nopenclaw dashboard`}
+          language="bash"
+        />
+      {:else if activeTab === "Pi"}
+        {#if runningModels.length > 1}
+          <div class="text-xs">
+            <span
+              class="text-exo-light-gray/50 text-[10px] uppercase tracking-wider block mb-1"
+              >Model</span
+            >
+            <select bind:value={piModel} class={selectClass}>
+              {#each runningModels as model}
+                <option value={model}>{model.split("/").pop()}</option>
+              {/each}
+            </select>
+          </div>
+        {/if}
+        <IntegrationCard
+          title="Models Config"
+          subtitle="~/.pi/agent/models.json"
+          description="Register exo as a custom provider in pi. Create or edit this file, then run pi and pick an exo model via /model. Install pi with: npm install -g @mariozechner/pi-coding-agent"
+          config={piModelsJson}
+        />
+        <IntegrationCard
+          title="Shell Command"
+          subtitle="Run in terminal"
+          description="Launch pi directly with the exo provider and model selected."
+          config={piShellCommand}
+          language="bash"
+        />
+      {:else if activeTab === "Open WebUI"}
+        <IntegrationCard
+          title="1. Start Open WebUI"
+          subtitle="Run in terminal"
+          description="Run this to start Open WebUI."
+          config={openWebUiCommand}
+          language="bash"
+        />
+        <IntegrationCard
+          title="2. Open & Select Model"
+          subtitle="http://localhost:3000"
+          description={`Open http://localhost:3000 in your browser. Select the running model from the dropdown at the top: ${runningModels.length > 0 ? runningModels.join(", ") : "no models running"}`}
+          config={"open http://localhost:3000"}
+          language="bash"
+        />
+        <IntegrationCard
+          title="Ollama CLI"
+          subtitle="Run in terminal"
+          description="Or use the Ollama CLI directly."
+          config={ollamaCommand}
+          language="bash"
+        />
+      {:else if activeTab === "n8n"}
+        <IntegrationCard
+          title="1. Start n8n"
+          subtitle="Run in terminal"
+          description="Start n8n with Docker. If you already have n8n running, skip this step."
+          config={n8nDockerCommand}
+          language="bash"
+        />
+        <IntegrationCard
+          title="2. Open n8n"
+          subtitle="http://localhost:5678"
+          description="Open n8n in your browser. If this is your first time, complete the setup and select 'Start from Scratch' when prompted."
+          config={"open http://localhost:5678"}
+          language="bash"
+        />
+        <IntegrationCard
+          title="3. Add OpenAI Credential"
+          subtitle="n8n UI → Credentials"
+          description="Create an OpenAI credential pointing at your exo cluster."
+          config={n8nCredentialSteps}
+        />
+        <IntegrationCard
+          title="4. Build a Workflow"
+          subtitle="n8n UI → Workflows"
+          description="Create a workflow that uses your exo-powered model."
+          config={n8nWorkflowSteps}
+        />
+      {:else if activeTab === "Firefox"}
+        <IntegrationCard
+          title="Firefox AI Chatbot"
+          subtitle="about:config"
+          description="Use the exo dashboard as Firefox's built-in AI chatbot. Requires Firefox 130+."
+          config={firefoxConfig}
+        />
+      {/if}
+    </div>
+  </main>
+</div>
@@ -81,4 +81,4 @@ Whenever a device produces side effects, it captures those side effects in an `E

 ## Purity

-A significant goal of the current design is to make data flow explicit. Classes should either represent simple data (`CamelCaseModel`s typically, and `TaggedModel`s for unions) or active `System`s (Erlang `Actor`s), with all transformations of that data being "referentially transparent" - destructure and construct new data, don't mutate in place. We have had varying degrees of success with this, and are still exploring where purity makes sense.
+A significant goal of the current design is to make data flow explicit. Classes should either represent simple data (`FrozenModel`s typically, and `TaggedModel`s for unions) or active `System`s (Erlang `Actor`s), with all transformations of that data being "referentially transparent" - destructure and construct new data, don't mutate in place. We have had varying degrees of success with this, and are still exploring where purity makes sense.
@@ -2,11 +2,11 @@
  "nodes": {
    "crane": {
      "locked": {
-        "lastModified": 1767744144,
-        "narHash": "sha256-9/9ntI0D+HbN4G0TrK3KmHbTvwgswz7p8IEJsWyef8Q=",
+        "lastModified": 1775790182,
+        "narHash": "sha256-pG2RWVQY0Pe+rmmXJx+Jpyi+JcgjWzS18m7fcD1B64Q=",
        "owner": "ipetkov",
        "repo": "crane",
-        "rev": "2fb033290bf6b23f226d4c8b32f7f7a16b043d7e",
+        "rev": "534982f1c41834b101e381b07b1121a4f065a374",
        "type": "github"
      },
      "original": {
@@ -47,11 +47,11 @@
        "rust-analyzer-src": "rust-analyzer-src"
      },
      "locked": {
-        "lastModified": 1768287139,
-        "narHash": "sha256-nsXFt0OzUi6K7dUzzJD5/v9e0Ic+fvclfIW936/43ZM=",
+        "lastModified": 1775807984,
+        "narHash": "sha256-Redoe3D9zGN5I9QPHWL9vfMVQBehY1fKsMiRXQ83X3w=",
        "owner": "nix-community",
        "repo": "fenix",
-        "rev": "a4a3aa956931f90f35453cb519e4545e9ad7f773",
+        "rev": "fcf90c0c4d368b2ca917a7afa6d08e98a397e5fd",
        "type": "github"
      },
      "original": {
@@ -83,11 +83,11 @@
        ]
      },
      "locked": {
-        "lastModified": 1768135262,
-        "narHash": "sha256-PVvu7OqHBGWN16zSi6tEmPwwHQ4rLPU9Plvs8/1TUBY=",
+        "lastModified": 1775087534,
+        "narHash": "sha256-91qqW8lhL7TLwgQWijoGBbiD4t7/q75KTi8NxjVmSmA=",
        "owner": "hercules-ci",
        "repo": "flake-parts",
-        "rev": "80daad04eddbbf5a4d883996a73f3f542fa437ac",
+        "rev": "3107b77cd68437b9a76194f0f7f9c55f2329ca5b",
        "type": "github"
      },
      "original": {
@@ -96,38 +96,42 @@
        "type": "github"
      }
    },
+    "nixglhost": {
+      "inputs": {
+        "nixpkgs": [
+          "nixpkgs"
+        ]
+      },
+      "locked": {
+        "lastModified": 1732211616,
+        "narHash": "sha256-QZCKJoypcwgS3tDNSWMjlxEBZtOYPW3eXV24rMzKsac=",
+        "owner": "numtide",
+        "repo": "nix-gl-host",
+        "rev": "5269b233f83880a0b433eafe026f0bc0d8f1a4a9",
+        "type": "github"
+      },
+      "original": {
+        "owner": "numtide",
+        "repo": "nix-gl-host",
+        "type": "github"
+      }
+    },
    "nixpkgs": {
      "locked": {
-        "lastModified": 1768127708,
-        "narHash": "sha256-1Sm77VfZh3mU0F5OqKABNLWxOuDeHIlcFjsXeeiPazs=",
+        "lastModified": 1775595990,
+        "narHash": "sha256-OEf7YqhF9IjJFYZJyuhAypgU+VsRB5lD4DuiMws5Ltc=",
        "owner": "NixOS",
        "repo": "nixpkgs",
-        "rev": "ffbc9f8cbaacfb331b6017d5a5abb21a492c9a38",
+        "rev": "4e92bbcdb030f3b4782be4751dc08e6b6cb6ccf2",
        "type": "github"
      },
      "original": {
        "owner": "NixOS",
-        "ref": "nixos-unstable",
+        "ref": "nixos-25.11",
        "repo": "nixpkgs",
        "type": "github"
      }
    },
-    "nixpkgs-swift": {
-      "locked": {
-        "lastModified": 1761672384,
-        "narHash": "sha256-o9KF3DJL7g7iYMZq9SWgfS1BFlNbsm6xplRjVlOCkXI=",
-        "owner": "NixOS",
-        "repo": "nixpkgs",
-        "rev": "08dacfca559e1d7da38f3cf05f1f45ee9bfd213c",
-        "type": "github"
-      },
-      "original": {
-        "owner": "NixOS",
-        "repo": "nixpkgs",
-        "rev": "08dacfca559e1d7da38f3cf05f1f45ee9bfd213c",
-        "type": "github"
-      }
-    },
    "purescript-overlay": {
      "inputs": {
        "flake-compat": "flake-compat",
@@ -164,11 +168,11 @@
        ]
      },
      "locked": {
-        "lastModified": 1763662255,
-        "narHash": "sha256-4bocaOyLa3AfiS8KrWjZQYu+IAta05u3gYZzZ6zXbT0=",
+        "lastModified": 1773870109,
+        "narHash": "sha256-ZoTdqZP03DcdoyxvpFHCAek4bkPUTUPUF3oCCgc3dP4=",
        "owner": "pyproject-nix",
        "repo": "build-system-pkgs",
-        "rev": "042904167604c681a090c07eb6967b4dd4dae88c",
+        "rev": "b6e74f433b02fa4b8a7965ee24680f4867e2926f",
        "type": "github"
      },
      "original": {
@@ -184,11 +188,11 @@
        ]
      },
      "locked": {
-        "lastModified": 1764134915,
-        "narHash": "sha256-xaKvtPx6YAnA3HQVp5LwyYG1MaN4LLehpQI8xEdBvBY=",
+        "lastModified": 1775439158,
+        "narHash": "sha256-NHY9SJNU019n+8NCabBDtmuzRFeE2gZlYKHowp9bV24=",
        "owner": "pyproject-nix",
        "repo": "pyproject.nix",
-        "rev": "2c8df1383b32e5443c921f61224b198a2282a657",
+        "rev": "fb6b728260f3f32761367e9fd1e1a25b4245bcd0",
        "type": "github"
      },
      "original": {
@@ -203,8 +207,8 @@
        "dream2nix": "dream2nix",
        "fenix": "fenix",
        "flake-parts": "flake-parts",
+        "nixglhost": "nixglhost",
        "nixpkgs": "nixpkgs",
-        "nixpkgs-swift": "nixpkgs-swift",
        "pyproject-build-systems": "pyproject-build-systems",
        "pyproject-nix": "pyproject-nix",
        "treefmt-nix": "treefmt-nix",
@@ -214,11 +218,11 @@
    "rust-analyzer-src": {
      "flake": false,
      "locked": {
-        "lastModified": 1768224240,
-        "narHash": "sha256-Pp1dDrXKPBUJReZnnDElFyHYn67XTd48zRhToheLjtk=",
+        "lastModified": 1775745684,
+        "narHash": "sha256-8MbfLwd60FNa8dRFkjE+G3TT/x21G3Rsplm1bMBQUtU=",
        "owner": "rust-lang",
        "repo": "rust-analyzer",
-        "rev": "725349602e525df37f377701e001fe8aab807878",
+        "rev": "64ddb549bc9a70d011328746fa46a8883f937b6b",
        "type": "github"
      },
      "original": {
@@ -257,11 +261,11 @@
        ]
      },
      "locked": {
-        "lastModified": 1768158989,
-        "narHash": "sha256-67vyT1+xClLldnumAzCTBvU0jLZ1YBcf4vANRWP3+Ak=",
+        "lastModified": 1775636079,
+        "narHash": "sha256-pc20NRoMdiar8oPQceQT47UUZMBTiMdUuWrYu2obUP0=",
        "owner": "numtide",
        "repo": "treefmt-nix",
-        "rev": "e96d59dff5c0d7fddb9d113ba108f03c3ef99eca",
+        "rev": "790751ff7fd3801feeaf96d7dc416a8d581265ba",
        "type": "github"
      },
      "original": {
@@ -280,11 +284,11 @@
        ]
      },
      "locked": {
-        "lastModified": 1767701098,
-        "narHash": "sha256-CJhKZnWb3gumR9oTRjFvCg/6lYTGbZRU7xtvcyWIRwU=",
+        "lastModified": 1775706324,
+        "narHash": "sha256-BTb4sydzX2B5/oNbvCdQFeSbk97xEnbb8bk84CiKCOs=",
        "owner": "pyproject-nix",
        "repo": "uv2nix",
-        "rev": "9d357f0d2ce6f5f35ec7959d7e704452352eb4da",
+        "rev": "5707df99097375896a3dda811d492a2fabe63500",
        "type": "github"
      },
      "original": {
@@ -2,7 +2,7 @@
  description = "The development environment for Exo";

  inputs = {
-    nixpkgs.url = "github:NixOS/nixpkgs/nixos-unstable";
+    nixpkgs.url = "github:NixOS/nixpkgs/nixos-25.11";

    flake-parts = {
      url = "github:hercules-ci/flake-parts";
@@ -46,17 +46,18 @@
      inputs.nixpkgs.follows = "nixpkgs";
    };

-    # Pinned nixpkgs for swift-format (swift is broken on x86_64-linux in newer nixpkgs)
-    nixpkgs-swift.url = "github:NixOS/nixpkgs/08dacfca559e1d7da38f3cf05f1f45ee9bfd213c";
+    nixglhost = {
+      url = "github:numtide/nix-gl-host";
+      inputs.nixpkgs.follows = "nixpkgs";
+    };
  };

  nixConfig = {
-    extra-trusted-public-keys = "exo.cachix.org-1:okq7hl624TBeAR3kV+g39dUFSiaZgLRkLsFBCuJ2NZI=";
-    extra-substituters = "https://exo.cachix.org";
+    extra-trusted-public-keys = "exo.cachix.org-1:okq7hl624TBeAR3kV+g39dUFSiaZgLRkLsFBCuJ2NZI= cache.nixos-cuda.org:74DUi4Ye579gUqzH4ziL9IyiJBlDpMRn9MBN8oNan9M=";
+    extra-substituters = "https://exo.cachix.org https://cache.nixos-cuda.org";
  };

-  outputs =
-    inputs:
+  outputs = inputs:
    inputs.flake-parts.lib.mkFlake { inherit inputs; } {
      systems = [
        "x86_64-linux"
@@ -71,32 +72,38 @@
        ./python/parts.nix
      ];

-      perSystem =
-        { config, self', pkgs, lib, system, ... }:
+      debug = true; # Enable options autocompletion
+
+      perSystem = { config, self', pkgs, lib, system, ... }:
        let
-          # Use pinned nixpkgs for swift-format (swift is broken on x86_64-linux in newer nixpkgs)
-          pkgsSwift = import inputs.nixpkgs-swift { inherit system; };
-        in
-        {
-          # Allow unfree for metal-toolchain (needed for Darwin Metal packages)
-          _module.args.pkgs = import inputs.nixpkgs {
+          pkgsArgs = {
            inherit system;
            config.allowUnfreePredicate = pkg: (pkg.pname or "") == "metal-toolchain";
            overlays = [
+              inputs.nixglhost.overlays.default
              (import ./nix/apple-sdk-overlay.nix)
-              (final: prev: {
-                macmon = prev.macmon.overrideAttrs (_: {
+              (final: _: {
+                macmon = final.rustPlatform.buildRustPackage {
+                  pname = "macmon";
                  version = "git";
                  src = final.fetchFromGitHub {
-                    owner = "swiftraccoon";
+                    owner = "vladkens";
                    repo = "macmon";
-                    rev = "9154d234f763fbeffdcb4135d0bbbaf80609699b";
-                    hash = "sha256-CwhilKNbs5XL9/tF5DMwyPBlE/hpmjGNTuxQ36sM50M=";
+                    rev = "a1cd06b6cc0d5e61db24fd8832e74cd992097a7d";
+                    hash = "sha256-wcq4PUXK44XfUKOZKl32u8LpOxXpSbUUfItQGwS2Zso=";
                  };
-                });
+                  cargoHash = "sha256-Epj3L+db1flGNK5y6yfSig8piEiXTz15lPo/FNkqlkA=";
+                };
              })
            ];
          };
+        in
+        {
+          # Allow unfree for metal-toolchain (needed for Darwin Metal packages)
+          _module.args = {
+            pkgs = import inputs.nixpkgs pkgsArgs;
+            unfreePkgs = import inputs.nixpkgs (pkgsArgs // { config.allowUnfree = true; });
+          };
          treefmt = {
            projectRootFile = "flake.nix";
            programs = {
@@ -116,29 +123,19 @@
              };
              swift-format = {
                enable = true;
-                package = pkgsSwift.swiftPackages.swift-format;
+                package = pkgs.swiftPackages.swift-format;
              };
              shfmt.enable = true;
              taplo.enable = true;
            };
          };

-          packages = lib.optionalAttrs pkgs.stdenv.hostPlatform.isDarwin (
-            let
-              uvLock = builtins.fromTOML (builtins.readFile ./uv.lock);
-              mlxPackage = builtins.head (builtins.filter (p: p.name == "mlx" && p.source ? git) uvLock.package);
-              uvLockMlxVersion = mlxPackage.version;
-              uvLockMlxRev = builtins.elemAt (builtins.split "#" mlxPackage.source.git) 2;
-            in
-            {
-              metal-toolchain = pkgs.callPackage ./nix/metal-toolchain.nix { };
-              mlx = pkgs.callPackage ./nix/mlx.nix {
-                inherit (self'.packages) metal-toolchain;
-                inherit uvLockMlxVersion uvLockMlxRev;
-              };
-              default = self'.packages.exo;
-            }
-          );
+          packages = {
+            default = self'.packages.exo;
+          } //
+          lib.optionalAttrs pkgs.stdenv.hostPlatform.isDarwin {
+            metal-toolchain = pkgs.callPackage ./nix/metal-toolchain.nix { };
+          };

          devShells.default = with pkgs; pkgs.mkShell {
            inputsFrom = [ self'.checks.cargo-build ];
@@ -149,16 +146,15 @@
                config.treefmt.build.wrapper

                # PYTHON
-                python313
+                self'.packages.editableVenv
                uv
-                ruff
-                basedpyright

                # RUST
                config.rust.toolchain
                maturin

                # NIX
+                nixd
                nixpkgs-fmt

                # SVELTE
@@ -168,9 +164,6 @@
                just
                jq
              ]
-              ++ lib.optionals stdenv.isLinux [
-                unixtools.ifconfig
-              ]
              ++ lib.optionals stdenv.isDarwin [
                macmon
              ];
@@ -22,19 +22,21 @@ sync-clean:
    uv sync --all-packages --force-reinstall --no-cache

 rust-rebuild:
-    cargo run --bin stub_gen
+    PYO3_PYTHON="$(uv run python -c 'import sys; print(sys.executable)')" cargo run --bin stub_gen
    uv sync --reinstall-package exo_pyo3_bindings

 build-dashboard:
    #!/usr/bin/env bash
-    cd dashboard
+    pushd dashboard
    npm install
    npm run build
+    popd

-package:
+package: build-dashboard
    uv run pyinstaller packaging/pyinstaller/exo.spec
+    rm -rf build

-build-app: package
+build-app: rust-rebuild sync-clean package
    xcodebuild build -project app/EXO/EXO.xcodeproj -scheme EXO -configuration Debug -derivedDataPath app/EXO/build
    @echo "\nBuild complete. Run with:\n  open {{justfile_directory()}}/app/EXO/build/Build/Products/Debug/EXO.app"

@@ -59,7 +59,7 @@ index 262b0495..5c7446ad 100644

 add_custom_command(
   OUTPUT ${MLX_METAL_PATH}/mlx.metallib
-  COMMAND xcrun -sdk macosx metallib ${KERNEL_AIR} -o
+-  COMMAND xcrun -sdk macosx metal ${KERNEL_AIR} -o
 +  COMMAND metallib ${KERNEL_AIR} -o
           ${MLX_METAL_PATH}/mlx.metallib
   DEPENDS ${KERNEL_AIR}
@@ -1,158 +0,0 @@
-{ stdenv
-, lib
-, fetchFromGitHub
-, replaceVars
-, fetchzip
-, cmake
-, nlohmann_json
-, apple-sdk_26
-, metal-toolchain
-, runCommand
-, fmt
-, python313Packages
-, uvLockMlxVersion
-, uvLockMlxRev
-}:
-
-assert stdenv.isDarwin;
-
-let
-  python = python313Packages.python;
-
-  # Static dependencies included directly during compilation
-  gguf-tools = fetchFromGitHub {
-    owner = "antirez";
-    repo = "gguf-tools";
-    rev = "8fa6eb65236618e28fd7710a0fba565f7faa1848";
-    hash = "sha256-15FvyPOFqTOr5vdWQoPnZz+mYH919++EtghjozDlnSA=";
-  };
-
-  metal_cpp = fetchzip {
-    url = "https://developer.apple.com/metal/cpp/files/metal-cpp_26.zip";
-    hash = "sha256-7n2eI2lw/S+Us6l7YPAATKwcIbRRpaQ8VmES7S8ZjY8=";
-  };
-
-  nanobind = fetchFromGitHub {
-    owner = "wjakob";
-    repo = "nanobind";
-    rev = "v2.10.2";
-    hash = "sha256-io44YhN+VpfHFWyvvLWSanRgbzA0whK8WlDNRi3hahU=";
-    fetchSubmodules = true;
-  };
-
-  mlx = stdenv.mkDerivation rec {
-    pname = "mlx";
-    version = uvLockMlxVersion;
-    pyproject = true;
-
-    src = fetchFromGitHub {
-      owner = "rltakashige";
-      repo = "mlx-jaccl-fix-small-recv";
-      rev = uvLockMlxRev;
-      hash = "sha256-GosFIWxIB48Egb1MqJrR3xhsUsQeWdRk5rV93USY6wQ=";
-    };
-
-    patches = [
-      (replaceVars ./darwin-build-fixes.patch {
-        sdkVersion = apple-sdk_26.version;
-        metalVersion = metal-toolchain.metalVersion;
-      })
-    ];
-
-    postPatch = ''
-      substituteInPlace mlx/backend/cpu/jit_compiler.cpp \
-        --replace-fail "g++" "$CXX"
-    '';
-
-    dontUseCmakeConfigure = true;
-
-    enableParallelBuilding = true;
-
-    # Allows multiple cores to be used in Python builds.
-    postUnpack = ''
-      export MAKEFLAGS+="''${enableParallelBuilding:+-j$NIX_BUILD_CORES}"
-    '';
-
-    # Updates the wrong fetcher rev attribute
-    passthru.skipBulkUpdate = true;
-
-    env = {
-      DEV_RELEASE = 1;
-      CMAKE_ARGS = toString [
-        (lib.cmakeBool "USE_SYSTEM_FMT" true)
-        (lib.cmakeOptionType "filepath" "FETCHCONTENT_SOURCE_DIR_GGUFLIB" "${gguf-tools}")
-        (lib.cmakeOptionType "filepath" "FETCHCONTENT_SOURCE_DIR_JSON" "${nlohmann_json.src}")
-        (lib.cmakeOptionType "filepath" "FETCHCONTENT_SOURCE_DIR_NANOBIND" "${nanobind}")
-        (lib.cmakeBool "FETCHCONTENT_FULLY_DISCONNECTED" true)
-        (lib.cmakeBool "MLX_BUILD_CPU" true)
-        (lib.cmakeBool "MLX_BUILD_METAL" true)
-        (lib.cmakeOptionType "filepath" "FETCHCONTENT_SOURCE_DIR_METAL_CPP" "${metal_cpp}")
-        (lib.cmakeOptionType "string" "CMAKE_OSX_DEPLOYMENT_TARGET" "${apple-sdk_26.version}")
-        (lib.cmakeOptionType "filepath" "CMAKE_OSX_SYSROOT" "${apple-sdk_26.passthru.sdkroot}")
-      ];
-      SDKROOT = apple-sdk_26.passthru.sdkroot;
-      MACOSX_DEPLOYMENT_TARGET = apple-sdk_26.version;
-    };
-
-    build-system = [
-      python313Packages.setuptools
-    ];
-
-    nativeBuildInputs = [
-      cmake
-      metal-toolchain
-      python313Packages.pypaBuildHook
-      python313Packages.pypaInstallHook
-      python313Packages.setuptools
-      python313Packages.typing-extensions
-      python313Packages.wheel
-      python313Packages.cmake
-      python313Packages.ninja
-    ];
-
-    buildInputs = [
-      fmt
-      gguf-tools
-      python313Packages.nanobind
-      python313Packages.pybind11
-      apple-sdk_26
-    ];
-
-    # Tests require Metal GPU access which isn't available in the Nix sandbox.
-    # To run tests, build with: nix build --option sandbox false .#mlx.passthru.tests.mlxTest
-    doCheck = false;
-
-    pythonImportsCheck = [ "mlx" ];
-
-    passthru.tests = {
-      # Runs example scripts to verify MLX works. Requires --option sandbox false
-      # since Metal GPU access is needed.
-      mlxTest =
-        runCommand "run-mlx-examples"
-          {
-            buildInputs = [ mlx ];
-            nativeBuildInputs = [ python ];
-          }
-          ''
-            cp ${src}/examples/python/logistic_regression.py .
-            ${python.interpreter} logistic_regression.py
-            rm logistic_regression.py
-
-            cp ${src}/examples/python/linear_regression.py .
-            ${python.interpreter} linear_regression.py
-            rm linear_regression.py
-
-            touch $out
-          '';
-    };
-
-    meta = {
-      homepage = "https://github.com/ml-explore/mlx";
-      description = "Array framework for Apple silicon";
-      changelog = "https://github.com/ml-explore/mlx/releases/tag/${src.tag}";
-      license = lib.licenses.mit;
-      platforms = [ "aarch64-darwin" ];
-    };
-  };
-in
-mlx
@@ -0,0 +1,6 @@
+{
+  "name": "exo",
+  "lockfileVersion": 3,
+  "requires": true,
+  "packages": {}
+}
@@ -1,5 +1,6 @@
 # -*- mode: python ; coding: utf-8 -*-

+import sys
 import importlib.util
 import shutil
 from pathlib import Path
@@ -56,6 +57,7 @@ HIDDEN_IMPORTS = sorted(
    set(
        collect_submodules("mlx")
        + _safe_collect("mlx_lm")
+        + _safe_collect("mlx_vlm")
        + _safe_collect("transformers")
    )
 )
@@ -67,18 +69,19 @@ DATAS: list[tuple[str, str]] = [
    (str(EXO_SHARED_MODELS_DIR), "exo/shared/models"),
 ]

-MACMON_PATH = shutil.which("macmon")
-if MACMON_PATH is None:
-    raise SystemExit(
-        "macmon binary not found in PATH. "
-        "Install the pinned fork used by exo via: "
-        "cargo install --git https://github.com/swiftraccoon/macmon "
-        "--rev 9154d234f763fbeffdcb4135d0bbbaf80609699b macmon --force"
-    )
+if sys.platform == "darwin":
+    MACMON_PATH = shutil.which("macmon")
+    if MACMON_PATH is None:
+        raise SystemExit(
+            "macmon binary not found in PATH. "
+            "Install the pinned fork used by exo via: "
+            "cargo install --git https://github.com/vladkens/macmon "
+            "--rev a1cd06b6cc0d5e61db24fd8832e74cd992097a7d macmon --force"
+        )

 BINARIES: list[tuple[str, str]] = [
    (MACMON_PATH, "."),
-]
+] if sys.platform == "darwin" else []

 a = Analysis(
    [str(ENTRYPOINT)],
@@ -1,9 +1,9 @@
 [project]
 name = "exo"
-version = "0.3.68"
+version = "0.3.70"
 description = "Exo"
 readme = "README.md"
-requires-python = ">=3.13"
+requires-python = "==3.13.*"
 dependencies = [
  "aiofiles>=24.1.0",
  "aiohttp>=3.12.14",
@@ -12,23 +12,24 @@ dependencies = [
  "fastapi>=0.116.1",
  "filelock>=3.18.0",
  "rustworkx>=0.17.1",
-  "huggingface-hub>=0.33.4",
+  "huggingface-hub>=1.8.0",
  "psutil>=7.0.0",
  "loguru>=0.7.3",
-  "exo_pyo3_bindings",                         # rust bindings
+  "exo-pyo3-bindings",                         # rust bindings
  "anyio==4.11.0",
-  "mlx; sys_platform == 'darwin'",
-  "mlx[cpu]==0.30.6; sys_platform == 'linux'",
-  "mlx-lm",
+  "mlx==0.31.2; sys_platform == 'darwin'",
+  "mlx-lm; sys_platform=='darwin'",
  "tiktoken>=0.12.0",                          # required for kimi k2 tokenizer
  "hypercorn>=0.18.0",
  "openai-harmony>=0.0.8",
  "httpx>=0.28.1",
  "tomlkit>=0.14.0",
-  "mflux==0.17.2",
+  "mflux==0.17.2; sys_platform == 'darwin'",
  "python-multipart>=0.0.21",
  "msgspec>=0.19.0",
  "zstandard>=0.23.0",
+  "mlx-vlm>=0.3.11; sys_platform == 'darwin'",
+  "transformers>=5.6.2",
 ]

 [project.scripts]
@@ -45,11 +46,29 @@ dev = [
  "ruff>=0.11.13",
 ]

-# mlx[cuda] requires a newer version of mlx. the ideal on linux is: default to mlx[cpu] unless[cuda] specified.
 [project.optional-dependencies]
-# cuda = [
-#     "mlx[cuda]==0.26.3",
-# ]
+build = ["nanobind"]
+cpu = [
+  "mlx==0.31.1; sys_platform == 'linux'",
+  "mlx-cpu==0.31.1; sys_platform == 'linux'",
+  "mlx-lm; sys_platform == 'linux'",
+  "mlx-vlm>=0.3.11; sys_platform== 'linux'",
+  "torch>=2.10.0; sys_platform == 'linux'",
+]
+cuda12 = [
+  "mlx==0.31.1; sys_platform == 'linux'",
+  "mlx-cuda-12==0.31.1; sys_platform == 'linux'",
+  "mlx-lm; sys_platform == 'linux'",
+  "mlx-vlm>=0.3.11; sys_platform== 'linux'",
+  "torch>=2.10.0; sys_platform == 'linux'",
+]
+cuda13 = [
+  "mlx==0.31.1; sys_platform == 'linux'",
+  "mlx-cuda-13==0.31.1; sys_platform == 'linux'",
+  "mlx-lm; sys_platform == 'linux'",
+  "mlx-vlm>=0.3.11; sys_platform== 'linux'",
+  "torch>=2.10.0; sys_platform == 'linux'",
+]

 ###
 # workspace configuration
@@ -59,12 +78,30 @@ dev = [
 members = ["rust/exo_pyo3_bindings", "bench"]

 [tool.uv.sources]
-exo_pyo3_bindings = { workspace = true }
+exo-pyo3-bindings = { workspace = true }
 mlx = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git", branch = "address-rdma-gpu-locks", marker = "sys_platform == 'darwin'" }
-mlx-lm = { git = "https://github.com/rltakashige/mlx-lm", branch = "leo/fix-deepseek-v32-indexer" }
-# Uncomment to use local mlx/mlx-lm development versions:
-# mlx = { path = "/Users/Shared/mlx", editable=true }
-# mlx-lm = { path = "/Users/Shared/mlx-lm", editable=true }
+mlx-lm = { git = "https://github.com/rltakashige/mlx-lm", branch = "leo/deepseek-v4" }
+torch = [
+  { index = "pytorch-cu130", marker = "sys_platform == 'linux' and extra == 'cuda13' and extra != 'cpu' and extra != 'cuda12'" },
+  { index = "pytorch-cu120", marker = "sys_platform == 'linux' and extra == 'cuda12' and extra != 'cpu' and extra != 'cuda13'" },
+  { index = "pytorch-cpu", marker = "(extra != 'cuda12' and extra != 'cuda13' and sys_platform == 'linux') or sys_platform == 'darwin'" },
+]
+vllm = { git = "https://github.com/hmellor/vllm.git", branch = "transformers-v5" }
+
+[[tool.uv.index]]
+name = "pytorch-cu130"
+url = "https://download.pytorch.org/whl/cu130"
+explicit = true
+
+[[tool.uv.index]]
+name = "pytorch-cu120"
+url = "https://download.pytorch.org/whl/cu120"
+explicit = true
+
+[[tool.uv.index]]
+name = "pytorch-cpu"
+url = "https://download.pytorch.org/whl/cpu"
+explicit = true

 [build-system]
 requires = ["uv_build>=0.8.9,<0.9.0"]
@@ -75,7 +112,7 @@ build-backend = "uv_build"
 ###

 [tool.basedpyright]
-include = [".venv/lib/mlx", ".venv/lib/mlx_lm", "src", "bench"]
+include = ["src", "bench"]
 typeCheckingMode = "strict"
 failOnWarnings = true

@@ -102,9 +139,14 @@ exclude = [
 ]
 stubPath = ".mlx_typings"

+[[tool.basedpyright.executionEnvironments]]
+root = "src/exo/worker/engines/image"
+reportMissingModuleSource = false
+
 [[tool.basedpyright.executionEnvironments]]
 root = "src"

+
 ###
 # uv configuration
 ###
@@ -114,6 +156,50 @@ root = "src"
 required-version = ">=0.8.6"
 prerelease = "allow"
 environments = ["sys_platform == 'darwin'", "sys_platform == 'linux'"]
+conflicts = [[{ extra = "cuda12" }, { extra = "cuda13" }, { extra = "cpu" }]]
+constraint-dependencies = ["transformers>=5.6.2"]
+override-dependencies = [
+  "mlx==0.31.1; sys_platform=='linux'",
+  "mlx; sys_platform=='darwin'",
+]
+
+[tool.uv.extra-build-dependencies]
+miniaudio = ["setuptools", "cffi", "pycparser"]
+mlx = [
+  "setuptools",
+  "typing-extensions",
+  "nanobind",
+  "pybind11",
+  "wheel",
+  "cmake",
+  "ninja",
+]
+mlx-lm = ["setuptools"]
+xgrammar = [
+  "nanobind",
+  "setuptools",
+  "scikit-build-core",
+  "packaging",
+  "pathspec",
+]
+rouge-score = ["setuptools"]
+sacrebleu = ["setuptools"]
+sqlitedict = ["setuptools"]
+word2number = ["setuptools"]
+vllm = [
+  "setuptools",
+  "setuptools-scm",
+  "scikit-build-core",
+  "jinja2",
+  "wheel",
+  "markupsafe",
+  "typing-extensions",
+  "torch",
+]
+fastsafetensors = ["setuptools", "pybind11"]
+torch = ["typing-extensions"]
+torchvision = ["torch"]
+torchaudio = ["torch"]

 ###
 # ruff configuration
@@ -121,7 +207,6 @@ environments = ["sys_platform == 'darwin'", "sys_platform == 'linux'"]

 [tool.ruff]
 extend-exclude = [
-  "shared/protobufs/**",
  "*mlx_typings/**",
  "rust/exo_pyo3_bindings/**",
  "bench/vendor/**",
@@ -1,18 +1,36 @@
 { inputs, ... }:
-{
-  perSystem =
-    { config, self', pkgs, lib, system, ... }:
+let
+  # Load workspace from uv.lock
+  workspace = inputs.uv2nix.lib.workspace.loadWorkspace {
+    workspaceRoot = ../.;
+  };
+
+  mkPythonSet = { pkgs, lib, self', members }:
    let
-      # Load workspace from uv.lock
-      workspace = inputs.uv2nix.lib.workspace.loadWorkspace {
-        workspaceRoot = inputs.self;
-      };
-
-      # Create overlay from workspace
-      # Use wheels from PyPI for most packages; we override mlx with our pure Nix Metal build
-      overlay = workspace.mkPyprojectOverlay { sourcePreference = "wheel"; };
-
-      # Override overlay to inject Nix-built components
+      inherit (pkgs.stdenv.hostPlatform) isLinux isDarwin isx86_64;
+      inherit (pkgs.config) cudaSupport;
+      inherit (pkgs) cudaPackages;
+      cuda13Support = cudaSupport && cudaPackages.cudaMajorVersion == "13";
+      libmlx_source = if cuda13Support then "mlx-cuda-13" else if cudaSupport then "mlx-cuda-12" else "mlx-cpu";
+      python = pkgs.python313;
+      cudaLibs = with cudaPackages; [
+        cuda_cudart
+        cuda_cccl
+        cuda_cupti
+        cuda_nvrtc
+        cuda_nvtx
+        cudnn
+        libcufile
+        libcublas
+        libcufft
+        libcurand
+        libcusolver
+        libcusparse
+        libcusparse_lt
+        libnvjitlink
+        libnvshmem
+        nccl
+      ];
      exoOverlay = final: prev: {
        # Replace workspace exo_pyo3_bindings with Nix-built wheel.
        # Preserve passthru so mkVirtualEnv can resolve dependency groups.
@@ -32,126 +50,162 @@
          '';
        };
      };
+      buildSystemsOverlay = final: prev:
+        lib.optionalAttrs isDarwin
+          {
+            mlx = prev.mlx.overrideAttrs (old:
+              let
+                # Static dependencies included directly during compilation
+                gguf-tools = pkgs.fetchFromGitHub {
+                  owner = "antirez";
+                  repo = "gguf-tools";
+                  rev = "8fa6eb65236618e28fd7710a0fba565f7faa1848";
+                  hash = "sha256-15FvyPOFqTOr5vdWQoPnZz+mYH919++EtghjozDlnSA=";
+                };

-      python = pkgs.python313;
+                metal_cpp = pkgs.fetchzip {
+                  url = "https://developer.apple.com/metal/cpp/files/metal-cpp_26.zip";
+                  hash = "sha256-7n2eI2lw/S+Us6l7YPAATKwcIbRRpaQ8VmES7S8ZjY8=";
+                };

-      # Overlay to provide build systems and custom packages
-      buildSystemsOverlay = final: prev: {
-        # mlx-lm is a git dependency that needs setuptools
-        mlx-lm = prev.mlx-lm.overrideAttrs (old: {
-          nativeBuildInputs = (old.nativeBuildInputs or [ ]) ++ [
-            final.setuptools
-          ];
-        });
-        # rouge-score and sacrebleu don't declare setuptools as a build dependency
-        rouge-score = prev.rouge-score.overrideAttrs (old: {
-          nativeBuildInputs = (old.nativeBuildInputs or [ ]) ++ [
-            final.setuptools
-          ];
-        });
-        sacrebleu = prev.sacrebleu.overrideAttrs (old: {
-          nativeBuildInputs = (old.nativeBuildInputs or [ ]) ++ [
-            final.setuptools
-          ];
-        });
-        sqlitedict = prev.sqlitedict.overrideAttrs (old: {
-          nativeBuildInputs = (old.nativeBuildInputs or [ ]) ++ [
-            final.setuptools
-          ];
-        });
-        word2number = prev.word2number.overrideAttrs (old: {
-          nativeBuildInputs = (old.nativeBuildInputs or [ ]) ++ [
-            final.setuptools
-          ];
-        });
-      } // lib.optionalAttrs pkgs.stdenv.hostPlatform.isDarwin {
-        # Use our pure Nix-built MLX with Metal support (macOS only)
-        mlx = self'.packages.mlx;
+                nanobind = pkgs.fetchFromGitHub {
+                  owner = "wjakob";
+                  repo = "nanobind";
+                  rev = "v2.10.2";
+                  hash = "sha256-io44YhN+VpfHFWyvvLWSanRgbzA0whK8WlDNRi3hahU=";
+                  fetchSubmodules = true;
+                };
+              in
+              {
+                nativeBuildInputs = (old.nativeBuildInputs or [ ]) ++ [ pkgs.cmake self'.packages.metal-toolchain ];
+                # TODO: non-sdk_26 support
+                buildInputs = (old.buildInputs or [ ])
+                ++ [ gguf-tools pkgs.fmt pkgs.nlohmann_json pkgs.apple-sdk_26 ];
+                patches = [
+                  (pkgs.replaceVars ../nix/darwin-build-fixes.patch {
+                    sdkVersion = pkgs.apple-sdk_26.version;
+                    inherit (self'.packages.metal-toolchain) metalVersion;
+                  })
+                ];
+                postPatch = ''
+                  substituteInPlace mlx/backend/cpu/jit_compiler.cpp \
+                    --replace-fail "g++" "${lib.getExe' pkgs.stdenv.cc "c++"}"
+                '';
+
+                DEV_RELEASE = 1;
+                CMAKE_ARGS = toString ([
+                  (lib.cmakeBool "USE_SYSTEM_FMT" true)
+                  (lib.cmakeOptionType "filepath" "FETCHCONTENT_SOURCE_DIR_GGUFLIB" "${gguf-tools}")
+                  (lib.cmakeOptionType "filepath" "FETCHCONTENT_SOURCE_DIR_JSON" "${pkgs.nlohmann_json.src}")
+                  (lib.cmakeOptionType "filepath" "FETCHCONTENT_SOURCE_DIR_NANOBIND" "${nanobind}")
+                  (lib.cmakeBool "FETCHCONTENT_FULLY_DISCONNECTED" true)
+                  (lib.cmakeBool "MLX_BUILD_CPU" true)
+                  (lib.cmakeBool "MLX_BUILD_METAL" true)
+                  (lib.cmakeOptionType "string" "CMAKE_INSTALL_LIBDIR" "lib")
+                  (lib.cmakeOptionType "filepath" "FETCHCONTENT_SOURCE_DIR_METAL_CPP" "${metal_cpp}")
+                  (lib.cmakeOptionType "string" "CMAKE_OSX_DEPLOYMENT_TARGET" "${pkgs.apple-sdk_26.version}")
+                  (lib.cmakeOptionType "filepath" "CMAKE_OSX_SYSROOT" "${pkgs.apple-sdk_26.passthru.sdkroot}")
+                ] ++ lib.optionals (isDarwin && isx86_64) [
+                  (lib.cmakeBool "MLX_ENABLE_X64_MAC" true)
+                ]);
+                SDKROOT = pkgs.apple-sdk_26.passthru.sdkroot;
+                MACOSX_DEPLOYMENT_TARGET = pkgs.apple-sdk_26.version;
+              });
+          } // lib.optionalAttrs isLinux {
+          mlx = prev.mlx.overrideAttrs (old: {
+            buildInputs = old.buildInputs ++ lib.optionals cudaSupport cudaLibs;
+            autoPatchelfIgnoreMissingDeps = lib.optionals cudaSupport [ "libcuda.so.1" ];
+            postInstall = ''
+              cp -r "${final.${libmlx_source}}/${final.python.sitePackages}/mlx" "$out/${final.python.sitePackages}/mlx/"
+            '';
+          });
+        } // lib.optionalAttrs cudaSupport {
+          "${libmlx_source}" = prev."${libmlx_source}".overrideAttrs (old: {
+            buildInputs = old.buildInputs ++ cudaLibs;
+            autoPatchelfIgnoreMissingDeps = [ "libcuda.so.1" ];
+          });
+          nvidia-cufile = prev.nvidia-cufile.overrideAttrs (old: {
+            buildInputs = old.buildInputs ++ [ pkgs.rdma-core ];
+            autoPatchelfIgnoreMissingDeps = [ "libcuda.so.1" ];
+          });
+          nvidia-cusolver = prev.nvidia-cusolver.overrideAttrs (old: {
+            buildInputs = old.buildInputs ++ cudaLibs;
+            autoPatchelfIgnoreMissingDeps = [ "libcuda.so.1" ];
+          });
+          nvidia-nvshmem-cu13 = prev.nvidia-nvshmem-cu13.overrideAttrs (old: {
+            buildInputs = old.buildInputs ++ [ pkgs.rdma-core pkgs.pmix pkgs.libfabric pkgs.ucx pkgs.openmpi ];
+            autoPatchelfIgnoreMissingDeps = [ "libcuda.so.1" ];
+          });
+          nvidia-cusparse = prev.nvidia-cusparse.overrideAttrs (old: {
+            buildInputs = old.buildInputs ++ [ cudaLibs ];
+            autoPatchelfIgnoreMissingDeps = [ "libcuda.so.1" ];
+          });
+          torch = prev.torch.overrideAttrs (old: {
+            buildInputs = old.buildInputs ++ cudaLibs;
+            autoPatchelfIgnoreMissingDeps = [ "libcuda.so.1" ];
+          });
+        };
+      pyprojectOverlay = workspace.mkPyprojectOverlay {
+        sourcePreference = "wheel";
+        dependencies = members;
+      };
+      editableOverlay = workspace.mkEditablePyprojectOverlay {
+        # Use environment variable pointing to editable root directory
+        root = "$REPO_ROOT";
+        members = [ "exo" "exo-bench" ];
      };
-
-      # Additional overlay for Linux-specific fixes (type checking env).
-      # Native wheels have shared lib dependencies we don't need at type-check time.
-      linuxOverlay = final: prev:
-        let
-          ignoreMissing = drv: drv.overrideAttrs { autoPatchelfIgnoreMissingDeps = [ "*" ]; };
-          nvidiaPackages = lib.filterAttrs (name: _: lib.hasPrefix "nvidia-" name) prev;
-        in
-        lib.optionalAttrs pkgs.stdenv.hostPlatform.isLinux (
-          (lib.mapAttrs (_: ignoreMissing) nvidiaPackages) // {
-            mlx = ignoreMissing prev.mlx;
-            mlx-cuda-13 = prev.mlx-cuda-13.overrideAttrs (old: {
-              buildInputs = (old.buildInputs or [ ]) ++ [
-                final.nvidia-cublas
-                final.nvidia-cuda-nvrtc
-                final.nvidia-cudnn-cu13
-                final.nvidia-nccl-cu13
-              ];
-              preFixup = ''
-                addAutoPatchelfSearchPath ${final.nvidia-cublas}
-                addAutoPatchelfSearchPath ${final.nvidia-cuda-nvrtc}
-                addAutoPatchelfSearchPath ${final.nvidia-cudnn-cu13}
-                addAutoPatchelfSearchPath ${final.nvidia-nccl-cu13}
-              '';
-              autoPatchelfIgnoreMissingDeps = [ "libcuda.so.1" ];
-            });
-            torch = ignoreMissing prev.torch;
-            triton = ignoreMissing prev.triton;
-          }
-        );
-
      pythonSet = (pkgs.callPackage inputs.pyproject-nix.build.packages {
        inherit python;
      }).overrideScope (
        lib.composeManyExtensions [
          inputs.pyproject-build-systems.overlays.default
-          overlay
+          pyprojectOverlay
          exoOverlay
          buildSystemsOverlay
-          linuxOverlay
        ]
      );
-      # mlx-cpu and mlx-cuda-13 both ship mlx/ site-packages files; keep first.
-      # mlx-cpu/mlx-cuda-13 and nvidia-cudnn-cu12/cu13 ship overlapping files.
-      venvCollisionPaths = lib.optionals pkgs.stdenv.hostPlatform.isLinux [
-        "lib/python3.13/site-packages/mlx*"
-        "lib/python3.13/site-packages/nvidia*"
-      ];
-
-      # Exclude bench deps from main env (bench has its own benchVenv)
-      exoDeps = removeAttrs workspace.deps.default [ "exo-bench" ];
-
-      exoVenv = (pythonSet.mkVirtualEnv "exo-env" exoDeps).overrideAttrs {
-        venvIgnoreCollisions = venvCollisionPaths;
-      };
-
-      # Virtual environment with dev dependencies for testing
-      testVenv = (pythonSet.mkVirtualEnv "exo-test-env" (
-        exoDeps // {
-          exo = [ "dev" ]; # Include pytest, pytest-asyncio, pytest-env
-        }
-      )).overrideAttrs {
-        venvIgnoreCollisions = venvCollisionPaths;
-      };
-
-      mkPythonScript = name: path: pkgs.writeShellApplication {
+      venv = name: (pythonSet.mkVirtualEnv "${name}-env" members).overrideAttrs (_: { venvSkip = [ "lib/python${python.pythonVersion}/site-packages/mlx/share/cmake/*" ]; });
+      mkApp = cmd: name: pkgs.writeShellApplication {
        inherit name;
-        runtimeInputs = [ exoVenv ];
        runtimeEnv = {
          EXO_DASHBOARD_DIR = self'.packages.dashboard;
          EXO_RESOURCES_DIR = inputs.self + /resources;
        };
-        text = ''exec python ${path} "$@"'';
+        runtimeInputs = [
+          # mlx and mlx-cuda ship clashing cmake files - we dont need them at runtime anyway
+          (venv name)
+        ]
+        ++ lib.optionals isDarwin [ pkgs.macmon ];
+        text = "exec " + lib.optionalString cudaSupport "${lib.getExe pkgs.nix-gl-host} " + cmd;
      };
+    in
+    {
+      inherit venv;
+      editablePythonSet = pythonSet.overrideScope editableOverlay;
+      mkPythonScript = path: mkApp ''python ${path} "$@"'';
+      mkExo = mkApp ''exo "$@"'';
+    };
+in
+{
+  perSystem =
+    { self', pkgs, unfreePkgs, lib, ... }:
+    let
+      inherit (pkgs.stdenv.hostPlatform) isLinux;
+      inherit (mkPythonSet { inherit self' pkgs lib; members = { exo = [ "cpu" ]; }; }) editablePythonSet mkExo;

-      benchVenv = pythonSet.mkVirtualEnv "exo-bench-env" {
-        exo-bench = [ ];
+      # Virtual environment with dev dependencies for testing
+      testVenv = (mkPythonSet {
+        inherit self' pkgs lib; members = {
+        exo = [ "dev" "cpu" ]; # Include pytest, pytest-asyncio, pytest-env
      };
+      }).venv "exo-test";

-      mkBenchScript = name: path: pkgs.writeShellApplication {
-        inherit name;
-        runtimeInputs = [ benchVenv ];
-        text = ''exec python ${path} "$@"'';
+      mkBenchScript = (mkPythonSet {
+        inherit self' pkgs lib; members = {
+        exo = [ "cpu" ];
+        exo-bench = [ ]; # Include pytest, pytest-asyncio, pytest-env
      };
+      }).mkPythonScript;

      mkSimplePythonScript = name: path: pkgs.writeShellApplication {
        inherit name;
@@ -159,56 +213,35 @@
        text = ''exec python ${path} "$@"'';
      };

-      exoPackage = pkgs.runCommand "exo"
-        {
-          nativeBuildInputs = [ pkgs.makeWrapper ];
-        }
-        ''
-          mkdir -p $out/bin
-
-          # Create wrapper script
-          makeWrapper ${exoVenv}/bin/exo $out/bin/exo \
-            --set EXO_DASHBOARD_DIR ${self'.packages.dashboard} \
-            --set EXO_RESOURCES_DIR ${inputs.self + /resources} \
-            ${lib.optionalString pkgs.stdenv.hostPlatform.isDarwin "--prefix PATH : ${pkgs.macmon}/bin"}
-        '';
    in
    {
-      # Python package only available on macOS (requires MLX/Metal)
-      packages = lib.optionalAttrs pkgs.stdenv.hostPlatform.isDarwin
-        {
-          exo = exoPackage;
-          # Test environment for running pytest outside of Nix sandbox (needs GPU access)
-          exo-test-env = testVenv;
-        } // {
+      packages = {
+        exo = mkExo "exo";
+        editableVenv = editablePythonSet.mkVirtualEnv "exo-dev-env" { exo = [ "dev" ]; };
+        # for running tests in ci
+        exo-test-env = testVenv;
        exo-bench = mkBenchScript "exo-bench" (inputs.self + /bench/exo_bench.py);
        exo-eval = mkBenchScript "exo-eval" (inputs.self + /bench/exo_eval.py);
        exo-eval-tool-calls = mkBenchScript "exo-eval-tool-calls" (inputs.self + /bench/eval_tool_calls.py);
+        # used by ./tests/run_exo_on.sh
        exo-get-all-models-on-cluster = mkSimplePythonScript "exo-get-all-models-on-cluster" (inputs.self + /tests/get_all_models_on_cluster.py);
+      } // lib.optionalAttrs isLinux {
+        exo-cuda-12 = (mkPythonSet { inherit self' lib; inherit (unfreePkgs.pkgsCuda.cudaPackages_12) pkgs; members = { exo = [ "cuda12" ]; }; }).mkExo "exo-cuda-12";
+        exo-cuda-13 = (mkPythonSet { inherit self' lib; inherit (unfreePkgs.pkgsCuda.cudaPackages_13) pkgs; members = { exo = [ "cuda13" ]; }; }).mkExo "exo-cuda-13";
      };

      checks = {
-        # Ruff linting (works on all platforms)
        lint = pkgs.runCommand "ruff-lint" { } ''
          export RUFF_CACHE_DIR="$TMPDIR/ruff-cache"
          ${pkgs.ruff}/bin/ruff check ${inputs.self}
          touch $out
        '';

-        # Hermetic basedpyright type checking
-        typecheck = pkgs.runCommand "typecheck"
-          {
-            nativeBuildInputs = [
-              testVenv
-              pkgs.basedpyright
-            ];
-          }
-          ''
-            cd ${inputs.self}
-            export HOME=$TMPDIR
-            basedpyright --pythonpath ${testVenv}/bin/python
-            touch $out
-          '';
+        typecheck = pkgs.runCommand "typecheck" { nativeBuildInputs = [ testVenv ]; } ''
+          cd ${inputs.self}
+          basedpyright
+          touch $out
+        '';
      };
    };
 }
@@ -8,6 +8,15 @@ family = "deepseek"
 quantization = "4bit"
 base_model = "DeepSeek V3.1"
 capabilities = ["text", "thinking", "thinking_toggle"]
+reasoning_dialect = "post_last_user"
+
+context_length = 131072

 [storage_size]
 in_bytes = 405874409472
+
+# Source: https://huggingface.co/deepseek-ai/DeepSeek-V3.1/blob/main/generation_config.json
+# Source: https://huggingface.co/deepseek-ai/DeepSeek-V3.1/discussions/19
+[sampling_defaults]
+temperature = 0.6
+top_p = 0.95
@@ -8,6 +8,15 @@ family = "deepseek"
 quantization = "8bit"
 base_model = "DeepSeek V3.1"
 capabilities = ["text", "thinking", "thinking_toggle"]
+reasoning_dialect = "post_last_user"
+
+context_length = 131072

 [storage_size]
 in_bytes = 765577920512
+
+# Source: https://huggingface.co/deepseek-ai/DeepSeek-V3.1/blob/main/generation_config.json
+# Source: https://huggingface.co/deepseek-ai/DeepSeek-V3.1/discussions/19
+[sampling_defaults]
+temperature = 0.6
+top_p = 0.95
@@ -8,6 +8,15 @@ family = "deepseek"
 quantization = "4bit"
 base_model = "DeepSeek V3.2"
 capabilities = ["text", "thinking", "thinking_toggle"]
+reasoning_dialect = "tool_conditional"
+
+context_length = 131072

 [storage_size]
 in_bytes = 378086226621
+
+# Source: https://huggingface.co/deepseek-ai/DeepSeek-V3.2/blob/main/generation_config.json
+# Source: https://docs.vllm.ai/projects/recipes/en/latest/DeepSeek/DeepSeek-V3_2.html
+[sampling_defaults]
+temperature = 1.0
+top_p = 0.95
@@ -8,6 +8,15 @@ family = "deepseek"
 quantization = "8bit"
 base_model = "DeepSeek V3.2"
 capabilities = ["text", "thinking", "thinking_toggle"]
+reasoning_dialect = "tool_conditional"
+
+context_length = 131072

 [storage_size]
 in_bytes = 755957120916
+
+# Source: https://huggingface.co/deepseek-ai/DeepSeek-V3.2/blob/main/generation_config.json
+# Source: https://docs.vllm.ai/projects/recipes/en/latest/DeepSeek/DeepSeek-V3_2.html
+[sampling_defaults]
+temperature = 1.0
+top_p = 0.95
@@ -0,0 +1,21 @@
+model_id = "mlx-community/DeepSeek-V4-Flash"
+n_layers = 43
+hidden_size = 4096
+num_key_value_heads = 1
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "deepseek"
+quantization = "8bit"
+base_model = "DeepSeek V4 Flash"
+capabilities = ["text", "thinking", "thinking_toggle"]
+reasoning_dialect = "tool_conditional"
+
+context_length = 1048576
+
+[storage_size]
+in_bytes = 155095760030
+
+# Source: https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash
+[sampling_defaults]
+temperature = 1.0
+top_p = 1.0
@@ -0,0 +1,21 @@
+model_id = "mlx-community/DeepSeek-V4-Pro"
+n_layers = 61
+hidden_size = 7168
+num_key_value_heads = 1
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "deepseek"
+quantization = "8bit"
+base_model = "DeepSeek V4 Pro"
+capabilities = ["text", "thinking", "thinking_toggle"]
+reasoning_dialect = "tool_conditional"
+
+context_length = 1048576
+
+[storage_size]
+in_bytes = 849681803879
+
+# Source: https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro
+[sampling_defaults]
+temperature = 1.0
+top_p = 1.0
@@ -8,6 +8,13 @@ family = "glm"
 quantization = "8bit"
 base_model = "GLM 4.5 Air"
 capabilities = ["text", "thinking", "thinking_toggle"]
+reasoning_dialect = "post_last_user"
+context_length = 131072

 [storage_size]
 in_bytes = 122406567936
+
+# Source: https://docs.z.ai/api-reference/llm/chat-completion
+[sampling_defaults]
+temperature = 0.6
+top_p = 0.95
@@ -8,6 +8,13 @@ family = "glm"
 quantization = "bf16"
 base_model = "GLM 4.5 Air"
 capabilities = ["text", "thinking", "thinking_toggle"]
+reasoning_dialect = "post_last_user"
+context_length = 131072

 [storage_size]
 in_bytes = 229780750336
+
+# Source: https://docs.z.ai/api-reference/llm/chat-completion
+[sampling_defaults]
+temperature = 0.6
+top_p = 0.95
@@ -8,6 +8,15 @@ family = "glm"
 quantization = "4bit"
 base_model = "GLM 4.7"
 capabilities = ["text", "thinking", "thinking_toggle"]
+reasoning_dialect = "post_last_user"
+context_length = 202752

 [storage_size]
 in_bytes = 198556925568
+
+# Source: https://huggingface.co/zai-org/GLM-4.7
+# Source: https://unsloth.ai/docs/models/glm-4.7-flash
+# Source: https://docs.z.ai/api-reference/llm/chat-completion
+[sampling_defaults]
+temperature = 1.0
+top_p = 0.95
@@ -8,6 +8,15 @@ family = "glm"
 quantization = "6bit"
 base_model = "GLM 4.7"
 capabilities = ["text", "thinking", "thinking_toggle"]
+reasoning_dialect = "post_last_user"
+context_length = 202752

 [storage_size]
 in_bytes = 286737579648
+
+# Source: https://huggingface.co/zai-org/GLM-4.7
+# Source: https://unsloth.ai/docs/models/glm-4.7-flash
+# Source: https://docs.z.ai/api-reference/llm/chat-completion
+[sampling_defaults]
+temperature = 1.0
+top_p = 0.95
@@ -8,6 +8,15 @@ family = "glm"
 quantization = "8bit"
 base_model = "GLM 4.7"
 capabilities = ["text", "thinking", "thinking_toggle"]
+reasoning_dialect = "post_last_user"
+context_length = 202752

 [storage_size]
 in_bytes = 396963397248
+
+# Source: https://huggingface.co/zai-org/GLM-4.7
+# Source: https://unsloth.ai/docs/models/glm-4.7-flash
+# Source: https://docs.z.ai/api-reference/llm/chat-completion
+[sampling_defaults]
+temperature = 1.0
+top_p = 0.95
@@ -8,6 +8,15 @@ family = "glm"
 quantization = "4bit"
 base_model = "GLM 4.7 Flash"
 capabilities = ["text", "thinking", "thinking_toggle"]
+reasoning_dialect = "post_last_user"
+context_length = 202752

 [storage_size]
 in_bytes = 19327352832
+
+# Source: https://huggingface.co/zai-org/GLM-4.7-Flash
+# Source: https://unsloth.ai/docs/models/glm-4.7-flash
+# Source: https://docs.z.ai/api-reference/llm/chat-completion
+[sampling_defaults]
+temperature = 1.0
+top_p = 0.95
@@ -8,6 +8,15 @@ family = "glm"
 quantization = "5bit"
 base_model = "GLM 4.7 Flash"
 capabilities = ["text", "thinking", "thinking_toggle"]
+reasoning_dialect = "post_last_user"
+context_length = 202752

 [storage_size]
 in_bytes = 22548578304
+
+# Source: https://huggingface.co/zai-org/GLM-4.7-Flash
+# Source: https://unsloth.ai/docs/models/glm-4.7-flash
+# Source: https://docs.z.ai/api-reference/llm/chat-completion
+[sampling_defaults]
+temperature = 1.0
+top_p = 0.95
@@ -8,6 +8,15 @@ family = "glm"
 quantization = "6bit"
 base_model = "GLM 4.7 Flash"
 capabilities = ["text", "thinking", "thinking_toggle"]
+reasoning_dialect = "post_last_user"
+context_length = 202752

 [storage_size]
 in_bytes = 26843545600
+
+# Source: https://huggingface.co/zai-org/GLM-4.7-Flash
+# Source: https://unsloth.ai/docs/models/glm-4.7-flash
+# Source: https://docs.z.ai/api-reference/llm/chat-completion
+[sampling_defaults]
+temperature = 1.0
+top_p = 0.95
@@ -8,6 +8,15 @@ family = "glm"
 quantization = "8bit"
 base_model = "GLM 4.7 Flash"
 capabilities = ["text", "thinking", "thinking_toggle"]
+reasoning_dialect = "post_last_user"
+context_length = 202752

 [storage_size]
 in_bytes = 34359738368
+
+# Source: https://huggingface.co/zai-org/GLM-4.7-Flash
+# Source: https://unsloth.ai/docs/models/glm-4.7-flash
+# Source: https://docs.z.ai/api-reference/llm/chat-completion
+[sampling_defaults]
+temperature = 1.0
+top_p = 0.95
@@ -8,6 +8,14 @@ family = "glm"
 quantization = "8bit"
 base_model = "GLM-5"
 capabilities = ["text", "thinking"]
+reasoning_dialect = "post_last_user"
+context_length = 202752

 [storage_size]
 in_bytes = 790517400864
+
+# Source: https://huggingface.co/zai-org/GLM-5
+# Source: https://docs.z.ai/api-reference/llm/chat-completion
+[sampling_defaults]
+temperature = 1.0
+top_p = 0.95
@@ -8,6 +8,14 @@ family = "glm"
 quantization = "MXFP4-Q8"
 base_model = "GLM-5"
 capabilities = ["text", "thinking"]
+reasoning_dialect = "post_last_user"
+context_length = 202752

 [storage_size]
 in_bytes = 405478939008
+
+# Source: https://huggingface.co/zai-org/GLM-5
+# Source: https://docs.z.ai/api-reference/llm/chat-completion
+[sampling_defaults]
+temperature = 1.0
+top_p = 0.95
@@ -8,6 +8,14 @@ family = "glm"
 quantization = "bf16"
 base_model = "GLM-5"
 capabilities = ["text", "thinking"]
+reasoning_dialect = "post_last_user"
+context_length = 202752

 [storage_size]
 in_bytes = 1487822475264
+
+# Source: https://huggingface.co/zai-org/GLM-5
+# Source: https://docs.z.ai/api-reference/llm/chat-completion
+[sampling_defaults]
+temperature = 1.0
+top_p = 0.95
@@ -0,0 +1,21 @@
+model_id = "mlx-community/GLM-5.1-DQ4plus-q8"
+n_layers = 78
+hidden_size = 6144
+num_key_value_heads = 64
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "glm"
+quantization = "8bit"
+base_model = "GLM-5.1"
+capabilities = ["text", "thinking"]
+reasoning_dialect = "post_last_user"
+context_length = 202752
+
+[storage_size]
+in_bytes = 465173655552
+
+# Source: https://huggingface.co/zai-org/GLM-5.1
+# Source: https://docs.z.ai/api-reference/llm/chat-completion
+[sampling_defaults]
+temperature = 1.0
+top_p = 0.95