move runner ack after status update

## Motivation

Enable parallel classifier-free guidance (CFG) for Qwen image models.
CFG requires two forward passes (positive/negative prompts) - this
allows them to run on separate nodes simultaneously, reducing latency.

## Changes

  - Added uses_cfg flag to ModelCard to identify CFG-based models
- Extended PipelineShardMetadata with CFG topology fields (cfg_rank,
cfg_world_size, peer device info)
- Updated placement to create two CFG groups with reversed ordering
(places CFG peers as ring neighbors)
- Refactored DiffusionRunner to process CFG branches separately with
exchange at last pipeline stage
- Added get_cfg_branch_data() to PromptData for single-branch embeddings
  - Fixed seed handling in API for distributed consistency
  - Fixed image yield to only emit from CFG rank 0 at last stage
  - Increased num_sync_steps_factor from 0.125 to 0.25 for Qwen

## Why It Works

- 2 nodes + CFG: Both run all layers, process different CFG branches in
parallel
  - 4+ even nodes + CFG: Hybrid - 2 CFG groups × N/2 pipeline stages
  - Odd nodes or non-CFG: Falls back to pure pipeline parallelism
 
 Ring topology places CFG peers as neighbors to enable direct exchange.

## Test Plan

### Manual Testing

Verified performance gain for Qwen-Image for 2 node and 4 node cluster.
Non-CFG models still work

### Automated Testing
 
Added tests in test_placement_utils.py covering 2-node CFG parallel,
4-node hybrid, odd-node fallback, and non-CFG pipeline modes.

.github/workflows/pipeline.yml

@@ -142,4 +142,6 @@ jobs:
           # Run pytest outside sandbox (needs GPU access for MLX)
           export HOME="$RUNNER_TEMP"
           export EXO_TESTS=1
+          export EXO_DASHBOARD_DIR="$PWD/dashboard/" 
+          export EXO_RESOURCES_DIR="$PWD/resources" 
           $TEST_ENV/bin/python -m pytest src -m "not slow" --import-mode=importlib

.gitignore

@@ -28,3 +28,7 @@ target/
 dashboard/build/
 dashboard/node_modules/
 dashboard/.svelte-kit/
+
+# host config snapshots
+hosts_*.json
+.swp

.mlx_typings/mlx_lm/tokenizer_utils.pyi

@@ -108,6 +108,7 @@ class TokenizerWrapper:
     _tokenizer: PreTrainedTokenizerFast
     eos_token_id: int | None
     eos_token: str | None
+    eos_token_ids: list[int] | set[int] | None
     bos_token_id: int | None
     bos_token: str | None
     vocab_size: int
@@ -117,7 +118,7 @@ class TokenizerWrapper:
         self,
         tokenizer: Any,
         detokenizer_class: Any = ...,
-        eos_token_ids: list[int] | None = ...,
+        eos_token_ids: list[int] | set[int] | None = ...,
         chat_template: Any = ...,
         tool_parser: Any = ...,
         tool_call_start: str | None = ...,

app/EXO/EXO/Models/ClusterState.swift

@@ -293,7 +293,7 @@ struct ClusterTask {
     let modelName: String?
     let promptPreview: String?
     let errorMessage: String?
-    let parameters: ChatCompletionTaskParameters?
+    let parameters: TextGenerationTaskParameters?
 
     var sortPriority: Int {
         switch status {
@@ -330,12 +330,12 @@ struct ClusterTaskPayload: Decodable {
     let taskStatus: TaskStatus?
     let instanceId: String?
     let commandId: String?
-    let taskParams: ChatCompletionTaskParameters?
+    let taskParams: TextGenerationTaskParameters?
     let errorType: String?
     let errorMessage: String?
 }
 
-struct ChatCompletionTaskParameters: Decodable, Equatable {
+struct TextGenerationTaskParameters: Decodable, Equatable {
     let model: String?
     let messages: [ChatCompletionMessage]?
     let maxTokens: Int?
@@ -374,7 +374,7 @@ extension ClusterTask {
         guard let id = payload.taskId else { return nil }
         let status = payload.taskStatus ?? .unknown
         switch kindKey {
-        case "ChatCompletion":
+        case "TextGeneration":
             self.init(
                 id: id,
                 status: status,

app/EXO/EXO/ViewModels/InstanceViewModel.swift

@@ -216,7 +216,7 @@ struct InstanceTaskViewModel: Identifiable, Equatable {
     let promptPreview: String?
     let errorMessage: String?
     let subtitle: String?
-    let parameters: ChatCompletionTaskParameters?
+    let parameters: TextGenerationTaskParameters?
 
     var title: String {
         switch kind {

bench/exo_bench.py

@@ -5,10 +5,13 @@ from __future__ import annotations
 import argparse
 import contextlib
 import http.client
+import itertools
 import json
 import os
+import sys
 import time
 from collections.abc import Callable
+from pathlib import Path
 from statistics import mean
 from typing import Any
 from urllib.parse import urlencode
@@ -16,6 +19,84 @@ from urllib.parse import urlencode
 from loguru import logger
 from transformers import AutoTokenizer
 
+# Monkey-patch for transformers 5.x compatibility
+# Kimi's tokenization_kimi.py imports bytes_to_unicode from the old location
+# which was moved in transformers 5.0.0rc2
+try:
+    import transformers.models.gpt2.tokenization_gpt2 as gpt2_tokenization
+    from transformers.convert_slow_tokenizer import bytes_to_unicode
+
+    if not hasattr(gpt2_tokenization, "bytes_to_unicode"):
+        gpt2_tokenization.bytes_to_unicode = bytes_to_unicode  # type: ignore[attr-defined]
+except ImportError:
+    pass  # transformers < 5.0 or bytes_to_unicode not available
+
+
+def load_tokenizer_for_bench(model_id: str) -> Any:
+    """
+    Load tokenizer for benchmarking, with special handling for Kimi models.
+
+    Kimi uses a custom TikTokenTokenizer that transformers 5.x can't load via AutoTokenizer.
+    This function replicates the logic from utils_mlx.py for bench compatibility.
+    """
+    model_id_lower = model_id.lower()
+
+    if "kimi-k2" in model_id_lower:
+        import importlib.util
+        import types
+
+        from huggingface_hub import snapshot_download
+
+        # Download/get the model path
+        model_path = Path(
+            snapshot_download(
+                model_id,
+                allow_patterns=["*.json", "*.py", "*.tiktoken"],
+            )
+        )
+
+        sys.path.insert(0, str(model_path))
+
+        # Load tool_declaration_ts first (tokenization_kimi imports it with relative import)
+        tool_decl_path = model_path / "tool_declaration_ts.py"
+        if tool_decl_path.exists():
+            spec = importlib.util.spec_from_file_location(
+                "tool_declaration_ts", tool_decl_path
+            )
+            if spec and spec.loader:
+                tool_decl_module = importlib.util.module_from_spec(spec)
+                sys.modules["tool_declaration_ts"] = tool_decl_module
+                spec.loader.exec_module(tool_decl_module)
+
+        # Load tokenization_kimi with patched source (convert relative to absolute import)
+        tok_path = model_path / "tokenization_kimi.py"
+        source = tok_path.read_text()
+        source = source.replace("from .tool_declaration_ts", "from tool_declaration_ts")
+        spec = importlib.util.spec_from_file_location("tokenization_kimi", tok_path)
+        if spec:
+            tok_module = types.ModuleType("tokenization_kimi")
+            tok_module.__file__ = str(tok_path)
+            sys.modules["tokenization_kimi"] = tok_module
+            exec(compile(source, tok_path, "exec"), tok_module.__dict__)  # noqa: S102
+            TikTokenTokenizer = tok_module.TikTokenTokenizer  # noqa: N806
+        else:
+            from tokenization_kimi import TikTokenTokenizer  # type: ignore[import-not-found]  # noqa: I001
+
+        hf_tokenizer: Any = TikTokenTokenizer.from_pretrained(model_path)
+
+        # Patch encode to use internal tiktoken model directly
+        # transformers 5.x has a bug in the encode->pad path for slow tokenizers
+        def _patched_encode(text: str, **kwargs: object) -> list[int]:
+            # Pass allowed_special="all" to handle special tokens like <|im_user|>
+            return list(hf_tokenizer.model.encode(text, allowed_special="all"))
+
+        hf_tokenizer.encode = _patched_encode
+
+        return hf_tokenizer
+
+    # Default: use AutoTokenizer
+    return AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
+
 
 class ExoHttpError(RuntimeError):
     def __init__(self, status: int, reason: str, body_preview: str):
@@ -24,7 +105,7 @@ class ExoHttpError(RuntimeError):
 
 
 class ExoClient:
-    def __init__(self, host: str, port: int, timeout_s: float = 600.0):
+    def __init__(self, host: str, port: int, timeout_s: float = 7200.0):
         self.host = host
         self.port = port
         self.timeout_s = timeout_s
@@ -180,14 +261,7 @@ def parse_int_list(values: list[str]) -> list[int]:
             part = part.strip()
             if part:
                 items.append(int(part))
-
-    seen: set[int] = set()
-    out: list[int] = []
-    for x in items:
-        if x not in seen:
-            out.append(x)
-            seen.add(x)
-    return out
+    return items
 
 
 def resolve_model_short_id(client: ExoClient, model_arg: str) -> tuple[str, str]:
@@ -240,7 +314,11 @@ def run_one_completion(
 
     stats = out.get("generation_stats")
 
-    preview = (out.get("choices") or [{}])[0]["message"]["content"][:200]
+    # 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 ""
 
     return {
         "elapsed_s": elapsed,
@@ -277,12 +355,29 @@ class PromptSizer:
                 f"Target ({target}) is smaller than template overhead ({self.base_tokens})."
             )
 
-        content = ""
-        tok = self.count_fn(content)
+        # Estimate tokens per atom using a sample
+        sample_count = 100
+        sample_content = self.atom * sample_count
+        sample_tokens = self.count_fn(sample_content) - self.base_tokens
+        tokens_per_atom = sample_tokens / sample_count
 
-        while tok < target:
-            content += self.atom
-            tok = self.count_fn(content)
+        # Estimate starting point
+        needed_tokens = target - self.base_tokens
+        estimated_atoms = int(needed_tokens / tokens_per_atom)
+
+        # Binary search to find exact atom count
+        low, high = 0, estimated_atoms * 2 + 100
+        while low < high:
+            mid = (low + high) // 2
+            tok = self.count_fn(self.atom * mid)
+            if tok < target:
+                low = mid + 1
+            else:
+                high = mid
+
+        content = self.atom * low
+        tok = self.count_fn(content)
+        logger.info(f"{tok=}")
 
         if tok != target:
             raise RuntimeError(
@@ -348,7 +443,7 @@ def main() -> int:
         help="Warmup runs per placement (uses first pp/tg).",
     )
     ap.add_argument(
-        "--timeout", type=float, default=600.0, help="HTTP timeout (seconds)."
+        "--timeout", type=float, default=7200.0, help="HTTP timeout (seconds)."
     )
     ap.add_argument(
         "--json-out",
@@ -358,6 +453,11 @@ def main() -> int:
     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 (cartesian product) even when lists have equal length.",
+    )
     args = ap.parse_args()
 
     pp_list = parse_int_list(args.pp)
@@ -369,6 +469,15 @@ def main() -> int:
         logger.error("--repeat must be >= 1")
         return 2
 
+    # Log pairing mode
+    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)
     short_id, full_model_id = resolve_model_short_id(client, args.model)
 
@@ -377,10 +486,7 @@ def main() -> int:
     )
     previews = previews_resp.get("previews") or []
 
-    tokenizer = AutoTokenizer.from_pretrained(
-        full_model_id,
-        trust_remote_code=True,
-    )
+    tokenizer = load_tokenizer_for_bench(full_model_id)
     if tokenizer is None:
         raise RuntimeError("[exo-bench] tokenizer load failed")
 
@@ -486,60 +592,55 @@ def main() -> int:
                 )
                 logger.debug(f"  warmup {i + 1}/{args.warmup} done")
 
-            for pp in pp_list:
-                # if (
-                #     pp * n_nodes > 2048
-                #     and "ring" in instance_meta.lower()
-                #     and "tensor" in sharding.lower()
-                # ):
-                #     model_card = MODEL_CARDS[short_id]
-                #     if model_card.metadata.storage_size > Memory.from_gb(10):
-                #         logger.info(
-                #             f"Skipping tensor ring as this is too slow for model of size {model_card.metadata.storage_size} on {n_nodes=}"
-                #         )
-                #         continue
-                for tg in tg_list:
-                    runs: list[dict[str, Any]] = []
-                    for r in range(args.repeat):
-                        time.sleep(3)
-                        try:
-                            row, actual_pp_tokens = run_one_completion(
-                                client, full_model_id, pp, tg, prompt_sizer
-                            )
-                        except Exception as e:
-                            logger.error(e)
-                            continue
-                        row.update(
-                            {
-                                "model_short_id": short_id,
-                                "model_id": full_model_id,
-                                "placement_sharding": sharding,
-                                "placement_instance_meta": instance_meta,
-                                "placement_nodes": n_nodes,
-                                "instance_id": instance_id,
-                                "pp_tokens": actual_pp_tokens,
-                                "tg": tg,
-                                "repeat_index": r,
-                            }
-                        )
-                        runs.append(row)
-                        all_rows.append(row)
+            # If pp and tg lists have same length, run in tandem (zip)
+            # Otherwise (or if --all-combinations), run all combinations (cartesian product)
+            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))
 
-                    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
+            for pp, tg in pp_tg_pairs:
+                runs: list[dict[str, Any]] = []
+                for r in range(args.repeat):
+                    time.sleep(3)
+                    try:
+                        row, actual_pp_tokens = run_one_completion(
+                            client, full_model_id, pp, tg, prompt_sizer
                         )
+                    except Exception as e:
+                        logger.error(e)
+                        continue
+                    row.update(
+                        {
+                            "model_short_id": short_id,
+                            "model_id": full_model_id,
+                            "placement_sharding": sharding,
+                            "placement_instance_meta": instance_meta,
+                            "placement_nodes": n_nodes,
+                            "instance_id": instance_id,
+                            "pp_tokens": actual_pp_tokens,
+                            "tg": tg,
+                            "repeat_index": r,
+                        }
+                    )
+                    runs.append(row)
+                    all_rows.append(row)
 
-                        logger.info(
-                            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"
-                        )
-                    time.sleep(2)
+                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
+                    )
+
+                    logger.info(
+                        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"
+                    )
+                time.sleep(2)
         finally:
             try:
                 client.request_json("DELETE", f"/instance/{instance_id}")

dashboard/src/lib/components/FamilyLogos.svelte

@@ -0,0 +1,73 @@
+<script lang="ts">
+  type FamilyLogoProps = {
+    family: string;
+    class?: string;
+  };
+
+  let { family, class: className = "" }: FamilyLogoProps = $props();
+</script>
+
+{#if family === "favorites"}
+  <svg class="w-6 h-6 {className}" viewBox="0 0 24 24" fill="currentColor">
+    <path
+      d="M12 2l3.09 6.26L22 9.27l-5 4.87 1.18 6.88L12 17.77l-6.18 3.25L7 14.14 2 9.27l6.91-1.01L12 2z"
+    />
+  </svg>
+{:else if family === "llama" || family === "meta"}
+  <svg class="w-6 h-6 {className}" viewBox="0 0 24 24" fill="currentColor">
+    <path
+      d="M6.915 4.03c-1.968 0-3.683 1.28-4.871 3.113C.704 9.208 0 11.883 0 14.449c0 .706.07 1.369.21 1.973a6.624 6.624 0 0 0 .265.86 5.297 5.297 0 0 0 .371.761c.696 1.159 1.818 1.927 3.593 1.927 1.497 0 2.633-.671 3.965-2.444.76-1.012 1.144-1.626 2.663-4.32l.756-1.339.186-.325c.061.1.121.196.183.3l2.152 3.595c.724 1.21 1.665 2.556 2.47 3.314 1.046.987 1.992 1.22 3.06 1.22 1.075 0 1.876-.355 2.455-.843a3.743 3.743 0 0 0 .81-.973c.542-.939.861-2.127.861-3.745 0-2.72-.681-5.357-2.084-7.45-1.282-1.912-2.957-2.93-4.716-2.93-1.047 0-2.088.467-3.053 1.308-.652.57-1.257 1.29-1.82 2.05-.69-.875-1.335-1.547-1.958-2.056-1.182-.966-2.315-1.303-3.454-1.303zm10.16 2.053c1.147 0 2.188.758 2.992 1.999 1.132 1.748 1.647 4.195 1.647 6.4 0 1.548-.368 2.9-1.839 2.9-.58 0-1.027-.23-1.664-1.004-.496-.601-1.343-1.878-2.832-4.358l-.617-1.028a44.908 44.908 0 0 0-1.255-1.98c.07-.109.141-.224.211-.327 1.12-1.667 2.118-2.602 3.358-2.602zm-10.201.553c1.265 0 2.058.791 2.675 1.446.307.327.737.871 1.234 1.579l-1.02 1.566c-.757 1.163-1.882 3.017-2.837 4.338-1.191 1.649-1.81 1.817-2.486 1.817-.524 0-1.038-.237-1.383-.794-.263-.426-.464-1.13-.464-2.046 0-2.221.63-4.535 1.66-6.088.454-.687.964-1.226 1.533-1.533a2.264 2.264 0 0 1 1.088-.285z"
+    />
+  </svg>
+{:else if family === "qwen"}
+  <svg class="w-6 h-6 {className}" viewBox="0 0 24 24" fill="currentColor">
+    <path
+      d="M12.604 1.34c.393.69.784 1.382 1.174 2.075a.18.18 0 00.157.091h5.552c.174 0 .322.11.446.327l1.454 2.57c.19.337.24.478.024.837-.26.43-.513.864-.76 1.3l-.367.658c-.106.196-.223.28-.04.512l2.652 4.637c.172.301.111.494-.043.77-.437.785-.882 1.564-1.335 2.34-.159.272-.352.375-.68.37-.777-.016-1.552-.01-2.327.016a.099.099 0 00-.081.05 575.097 575.097 0 01-2.705 4.74c-.169.293-.38.363-.725.364-.997.003-2.002.004-3.017.002a.537.537 0 01-.465-.271l-1.335-2.323a.09.09 0 00-.083-.049H4.982c-.285.03-.553-.001-.805-.092l-1.603-2.77a.543.543 0 01-.002-.54l1.207-2.12a.198.198 0 000-.197 550.951 550.951 0 01-1.875-3.272l-.79-1.395c-.16-.31-.173-.496.095-.965.465-.813.927-1.625 1.387-2.436.132-.234.304-.334.584-.335a338.3 338.3 0 012.589-.001.124.124 0 00.107-.063l2.806-4.895a.488.488 0 01.422-.246c.524-.001 1.053 0 1.583-.006L11.704 1c.341-.003.724.032.9.34zm-3.432.403a.06.06 0 00-.052.03L6.254 6.788a.157.157 0 01-.135.078H3.253c-.056 0-.07.025-.041.074l5.81 10.156c.025.042.013.062-.034.063l-2.795.015a.218.218 0 00-.2.116l-1.32 2.31c-.044.078-.021.118.068.118l5.716.008c.046 0 .08.02.104.061l1.403 2.454c.046.081.092.082.139 0l5.006-8.76.783-1.382a.055.055 0 01.096 0l1.424 2.53a.122.122 0 00.107.062l2.763-.02a.04.04 0 00.035-.02.041.041 0 000-.04l-2.9-5.086a.108.108 0 010-.113l.293-.507 1.12-1.977c.024-.041.012-.062-.035-.062H9.2c-.059 0-.073-.026-.043-.077l1.434-2.505a.107.107 0 000-.114L9.225 1.774a.06.06 0 00-.053-.031zm6.29 8.02c.046 0 .058.02.034.06l-.832 1.465-2.613 4.585a.056.056 0 01-.05.029.058.058 0 01-.05-.029L8.498 9.841c-.02-.034-.01-.052.028-.054l.216-.012 6.722-.012z"
+    />
+  </svg>
+{:else if family === "deepseek"}
+  <svg class="w-6 h-6 {className}" viewBox="0 0 24 24" fill="currentColor">
+    <path
+      d="M23.748 4.482c-.254-.124-.364.113-.512.234-.051.039-.094.09-.137.136-.372.397-.806.657-1.373.626-.829-.046-1.537.214-2.163.848-.133-.782-.575-1.248-1.247-1.548-.352-.156-.708-.311-.955-.65-.172-.241-.219-.51-.305-.774-.055-.16-.11-.323-.293-.35-.2-.031-.278.136-.356.276-.313.572-.434 1.202-.422 1.84.027 1.436.633 2.58 1.838 3.393.137.093.172.187.129.323-.082.28-.18.552-.266.833-.055.179-.137.217-.329.14a5.526 5.526 0 01-1.736-1.18c-.857-.828-1.631-1.742-2.597-2.458a11.365 11.365 0 00-.689-.471c-.985-.957.13-1.743.388-1.836.27-.098.093-.432-.779-.428-.872.004-1.67.295-2.687.684a3.055 3.055 0 01-.465.137 9.597 9.597 0 00-2.883-.102c-1.885.21-3.39 1.102-4.497 2.623C.082 8.606-.231 10.684.152 12.85c.403 2.284 1.569 4.175 3.36 5.653 1.858 1.533 3.997 2.284 6.438 2.14 1.482-.085 3.133-.284 4.994-1.86.47.234.962.327 1.78.397.63.059 1.236-.03 1.705-.128.735-.156.684-.837.419-.961-2.155-1.004-1.682-.595-2.113-.926 1.096-1.296 2.746-2.642 3.392-7.003.05-.347.007-.565 0-.845-.004-.17.035-.237.23-.256a4.173 4.173 0 001.545-.475c1.396-.763 1.96-2.015 2.093-3.517.02-.23-.004-.467-.247-.588zM11.581 18c-2.089-1.642-3.102-2.183-3.52-2.16-.392.024-.321.471-.235.763.09.288.207.486.371.739.114.167.192.416-.113.603-.673.416-1.842-.14-1.897-.167-1.361-.802-2.5-1.86-3.301-3.307-.774-1.393-1.224-2.887-1.298-4.482-.02-.386.093-.522.477-.592a4.696 4.696 0 011.529-.039c2.132.312 3.946 1.265 5.468 2.774.868.86 1.525 1.887 2.202 2.891.72 1.066 1.494 2.082 2.48 2.914.348.292.625.514.891.677-.802.09-2.14.11-3.054-.614zm1-6.44a.306.306 0 01.415-.287.302.302 0 01.2.288.306.306 0 01-.31.307.303.303 0 01-.304-.308zm3.11 1.596c-.2.081-.399.151-.59.16a1.245 1.245 0 01-.798-.254c-.274-.23-.47-.358-.552-.758a1.73 1.73 0 01.016-.588c.07-.327-.008-.537-.239-.727-.187-.156-.426-.199-.688-.199a.559.559 0 01-.254-.078c-.11-.054-.2-.19-.114-.358.028-.054.16-.186.192-.21.356-.202.767-.136 1.146.016.352.144.618.408 1.001.782.391.451.462.576.685.914.176.265.336.537.445.848.067.195-.019.354-.25.452z"
+    />
+  </svg>
+{:else if family === "openai" || family === "gpt-oss"}
+  <svg class="w-6 h-6 {className}" viewBox="0 0 24 24" fill="currentColor">
+    <path
+      d="M22.2819 9.8211a5.9847 5.9847 0 0 0-.5157-4.9108 6.0462 6.0462 0 0 0-6.5098-2.9A6.0651 6.0651 0 0 0 4.9807 4.1818a5.9847 5.9847 0 0 0-3.9977 2.9 6.0462 6.0462 0 0 0 .7427 7.0966 5.98 5.98 0 0 0 .511 4.9107 6.051 6.051 0 0 0 6.5146 2.9001A5.9847 5.9847 0 0 0 13.2599 24a6.0557 6.0557 0 0 0 5.7718-4.2058 5.9894 5.9894 0 0 0 3.9977-2.9001 6.0557 6.0557 0 0 0-.7475-7.0729zm-9.022 12.6081a4.4755 4.4755 0 0 1-2.8764-1.0408l.1419-.0804 4.7783-2.7582a.7948.7948 0 0 0 .3927-.6813v-6.7369l2.02 1.1686a.071.071 0 0 1 .038.052v5.5826a4.504 4.504 0 0 1-4.4945 4.4944zm-9.6607-4.1254a4.4708 4.4708 0 0 1-.5346-3.0137l.142.0852 4.783 2.7582a.7712.7712 0 0 0 .7806 0l5.8428-3.3685v2.3324a.0804.0804 0 0 1-.0332.0615L9.74 19.9502a4.4992 4.4992 0 0 1-6.1408-1.6464zM2.3408 7.8956a4.485 4.485 0 0 1 2.3655-1.9728V11.6a.7664.7664 0 0 0 .3879.6765l5.8144 3.3543-2.0201 1.1685a.0757.0757 0 0 1-.071 0l-4.8303-2.7865A4.504 4.504 0 0 1 2.3408 7.872zm16.5963 3.8558L13.1038 8.364 15.1192 7.2a.0757.0757 0 0 1 .071 0l4.8303 2.7913a4.4944 4.4944 0 0 1-.6765 8.1042v-5.6772a.79.79 0 0 0-.407-.667zm2.0107-3.0231l-.142-.0852-4.7735-2.7818a.7759.7759 0 0 0-.7854 0L9.409 9.2297V6.8974a.0662.0662 0 0 1 .0284-.0615l4.8303-2.7866a4.4992 4.4992 0 0 1 6.6802 4.66zM8.3065 12.863l-2.02-1.1638a.0804.0804 0 0 1-.038-.0567V6.0742a4.4992 4.4992 0 0 1 7.3757-3.4537l-.142.0805L8.704 5.459a.7948.7948 0 0 0-.3927.6813zm1.0976-2.3654l2.602-1.4998 2.6069 1.4998v2.9994l-2.5974 1.4997-2.6067-1.4997Z"
+    />
+  </svg>
+{:else if family === "glm"}
+  <svg class="w-6 h-6 {className}" viewBox="0 0 24 24" fill="currentColor">
+    <path
+      d="M11.991 23.503a.24.24 0 00-.244.248.24.24 0 00.244.249.24.24 0 00.245-.249.24.24 0 00-.22-.247l-.025-.001zM9.671 5.365a1.697 1.697 0 011.099 2.132l-.071.172-.016.04-.018.054c-.07.16-.104.32-.104.498-.035.71.47 1.279 1.186 1.314h.366c1.309.053 2.338 1.173 2.286 2.523-.052 1.332-1.152 2.38-2.478 2.327h-.174c-.715.018-1.274.64-1.239 1.368 0 .124.018.23.053.337.209.373.54.658.96.8.75.23 1.517-.125 1.9-.782l.018-.035c.402-.64 1.17-.96 1.92-.711.854.284 1.378 1.226 1.099 2.167a1.661 1.661 0 01-2.077 1.102 1.711 1.711 0 01-.907-.711l-.017-.035c-.2-.323-.463-.58-.851-.711l-.056-.018a1.646 1.646 0 00-1.954.746 1.66 1.66 0 01-1.065.764 1.677 1.677 0 01-1.989-1.279c-.209-.906.332-1.83 1.257-2.043a1.51 1.51 0 01.296-.035h.018c.68-.071 1.151-.622 1.116-1.333a1.307 1.307 0 00-.227-.693 2.515 2.515 0 01-.366-1.403 2.39 2.39 0 01.366-1.208c.14-.195.21-.444.227-.693.018-.71-.506-1.261-1.186-1.332l-.07-.018a1.43 1.43 0 01-.299-.07l-.05-.019a1.7 1.7 0 01-1.047-2.114 1.68 1.68 0 012.094-1.101zm-5.575 10.11c.26-.264.639-.367.994-.27.355.096.633.379.728.74.095.362-.007.748-.267 1.013-.402.41-1.053.41-1.455 0a1.062 1.062 0 010-1.482zm14.845-.294c.359-.09.738.024.992.297.254.274.344.665.237 1.025-.107.36-.396.634-.756.718-.551.128-1.1-.22-1.23-.781a1.05 1.05 0 01.757-1.26zm-.064-4.39c.314.32.49.753.49 1.206 0 .452-.176.886-.49 1.206-.315.32-.74.5-1.185.5-.444 0-.87-.18-1.184-.5a1.727 1.727 0 010-2.412 1.654 1.654 0 012.369 0zm-11.243.163c.364.484.447 1.128.218 1.691a1.665 1.665 0 01-2.188.923c-.855-.36-1.26-1.358-.907-2.228a1.68 1.68 0 011.33-1.038c.593-.08 1.183.169 1.547.652zm11.545-4.221c.368 0 .708.2.892.524.184.324.184.724 0 1.048a1.026 1.026 0 01-.892.524c-.568 0-1.03-.47-1.03-1.048 0-.579.462-1.048 1.03-1.048zm-14.358 0c.368 0 .707.2.891.524.184.324.184.724 0 1.048a1.026 1.026 0 01-.891.524c-.569 0-1.03-.47-1.03-1.048 0-.579.461-1.048 1.03-1.048zm10.031-1.475c.925 0 1.675.764 1.675 1.706s-.75 1.705-1.675 1.705-1.674-.763-1.674-1.705c0-.942.75-1.706 1.674-1.706zm-2.626-.684c.362-.082.653-.356.761-.718a1.062 1.062 0 00-.238-1.028 1.017 1.017 0 00-.996-.294c-.547.14-.881.7-.752 1.257.13.558.675.907 1.225.783zm0 16.876c.359-.087.644-.36.75-.72a1.062 1.062 0 00-.237-1.019 1.018 1.018 0 00-.985-.301 1.037 1.037 0 00-.762.717c-.108.361-.017.754.239 1.028.245.263.606.377.953.305l.043-.01zM17.19 3.5a.631.631 0 00.628-.64c0-.355-.279-.64-.628-.64a.631.631 0 00-.628.64c0 .355.28.64.628.64zm-10.38 0a.631.631 0 00.628-.64c0-.355-.28-.64-.628-.64a.631.631 0 00-.628.64c0 .355.279.64.628.64zm-5.182 7.852a.631.631 0 00-.628.64c0 .354.28.639.628.639a.63.63 0 00.627-.606l.001-.034a.62.62 0 00-.628-.64zm5.182 9.13a.631.631 0 00-.628.64c0 .355.279.64.628.64a.631.631 0 00.628-.64c0-.355-.28-.64-.628-.64zm10.38.018a.631.631 0 00-.628.64c0 .355.28.64.628.64a.631.631 0 00.628-.64c0-.355-.279-.64-.628-.64zm5.182-9.148a.631.631 0 00-.628.64c0 .354.279.639.628.639a.631.631 0 00.628-.64c0-.355-.28-.64-.628-.64zm-.384-4.992a.24.24 0 00.244-.249.24.24 0 00-.244-.249.24.24 0 00-.244.249c0 .142.122.249.244.249zM11.991.497a.24.24 0 00.245-.248A.24.24 0 0011.99 0a.24.24 0 00-.244.249c0 .133.108.236.223.247l.021.001zM2.011 6.36a.24.24 0 00.245-.249.24.24 0 00-.244-.249.24.24 0 00-.244.249.24.24 0 00.244.249zm0 11.263a.24.24 0 00-.243.248.24.24 0 00.244.249.24.24 0 00.244-.249.252.252 0 00-.244-.248zm19.995-.018a.24.24 0 00-.245.248.24.24 0 00.245.25.24.24 0 00.244-.25.252.252 0 00-.244-.248z"
+    />
+  </svg>
+{:else if family === "minimax"}
+  <svg class="w-6 h-6 {className}" viewBox="0 0 24 24" fill="currentColor">
+    <path
+      d="M16.278 2c1.156 0 2.093.927 2.093 2.07v12.501a.74.74 0 00.744.709.74.74 0 00.743-.709V9.099a2.06 2.06 0 012.071-2.049A2.06 2.06 0 0124 9.1v6.561a.649.649 0 01-.652.645.649.649 0 01-.653-.645V9.1a.762.762 0 00-.766-.758.762.762 0 00-.766.758v7.472a2.037 2.037 0 01-2.048 2.026 2.037 2.037 0 01-2.048-2.026v-12.5a.785.785 0 00-.788-.753.785.785 0 00-.789.752l-.001 15.904A2.037 2.037 0 0113.441 22a2.037 2.037 0 01-2.048-2.026V18.04c0-.356.292-.645.652-.645.36 0 .652.289.652.645v1.934c0 .263.142.506.372.638.23.131.514.131.744 0a.734.734 0 00.372-.638V4.07c0-1.143.937-2.07 2.093-2.07zm-5.674 0c1.156 0 2.093.927 2.093 2.07v11.523a.648.648 0 01-.652.645.648.648 0 01-.652-.645V4.07a.785.785 0 00-.789-.78.785.785 0 00-.789.78v14.013a2.06 2.06 0 01-2.07 2.048 2.06 2.06 0 01-2.071-2.048V9.1a.762.762 0 00-.766-.758.762.762 0 00-.766.758v3.8a2.06 2.06 0 01-2.071 2.049A2.06 2.06 0 010 12.9v-1.378c0-.357.292-.646.652-.646.36 0 .653.29.653.646V12.9c0 .418.343.757.766.757s.766-.339.766-.757V9.099a2.06 2.06 0 012.07-2.048 2.06 2.06 0 012.071 2.048v8.984c0 .419.343.758.767.758.423 0 .766-.339.766-.758V4.07c0-1.143.937-2.07 2.093-2.07z"
+    />
+  </svg>
+{:else if family === "kimi"}
+  <svg class="w-6 h-6 {className}" viewBox="0 0 24 24" fill="currentColor">
+    <path
+      d="M19.738 5.776c.163-.209.306-.4.457-.585.07-.087.064-.153-.004-.244-.655-.861-.717-1.817-.34-2.787.283-.73.909-1.072 1.674-1.145.477-.045.945.004 1.379.236.57.305.902.77 1.01 1.412.086.512.07 1.012-.075 1.508-.257.878-.888 1.333-1.753 1.448-.718.096-1.446.108-2.17.157-.056.004-.113 0-.178 0z"
+    />
+    <path
+      d="M17.962 1.844h-4.326l-3.425 7.81H5.369V1.878H1.5V22h3.87v-8.477h6.824a3.025 3.025 0 002.743-1.75V22h3.87v-8.477a3.87 3.87 0 00-3.588-3.86v-.01h-2.125a3.94 3.94 0 002.323-2.12l2.545-5.689z"
+    />
+  </svg>
+{:else if family === "huggingface"}
+  <svg class="w-6 h-6 {className}" viewBox="0 0 24 24" fill="currentColor">
+    <path
+      d="M12.025 1.13c-5.77 0-10.449 4.647-10.449 10.378 0 1.112.178 2.181.503 3.185.064-.222.203-.444.416-.577a.96.96 0 0 1 .524-.15c.293 0 .584.124.84.284.278.173.48.408.71.694.226.282.458.611.684.951v-.014c.017-.324.106-.622.264-.874s.403-.487.762-.543c.3-.047.596.06.787.203s.31.313.4.467c.15.257.212.468.233.542.01.026.653 1.552 1.657 2.54.616.605 1.01 1.223 1.082 1.912.055.537-.096 1.059-.38 1.572.637.121 1.294.187 1.967.187.657 0 1.298-.063 1.921-.178-.287-.517-.44-1.041-.384-1.581.07-.69.465-1.307 1.081-1.913 1.004-.987 1.647-2.513 1.657-2.539.021-.074.083-.285.233-.542.09-.154.208-.323.4-.467a1.08 1.08 0 0 1 .787-.203c.359.056.604.29.762.543s.247.55.265.874v.015c.225-.34.457-.67.683-.952.23-.286.432-.52.71-.694.257-.16.547-.284.84-.285a.97.97 0 0 1 .524.151c.228.143.373.388.43.625l.006.04a10.3 10.3 0 0 0 .534-3.273c0-5.731-4.678-10.378-10.449-10.378M8.327 6.583a1.5 1.5 0 0 1 .713.174 1.487 1.487 0 0 1 .617 2.013c-.183.343-.762-.214-1.102-.094-.38.134-.532.914-.917.71a1.487 1.487 0 0 1 .69-2.803m7.486 0a1.487 1.487 0 0 1 .689 2.803c-.385.204-.536-.576-.916-.71-.34-.12-.92.437-1.103.094a1.487 1.487 0 0 1 .617-2.013 1.5 1.5 0 0 1 .713-.174m-10.68 1.55a.96.96 0 1 1 0 1.921.96.96 0 0 1 0-1.92m13.838 0a.96.96 0 1 1 0 1.92.96.96 0 0 1 0-1.92M8.489 11.458c.588.01 1.965 1.157 3.572 1.164 1.607-.007 2.984-1.155 3.572-1.164.196-.003.305.12.305.454 0 .886-.424 2.328-1.563 3.202-.22-.756-1.396-1.366-1.63-1.32q-.011.001-.02.006l-.044.026-.01.008-.03.024q-.018.017-.035.036l-.032.04a1 1 0 0 0-.058.09l-.014.025q-.049.088-.11.19a1 1 0 0 1-.083.116 1.2 1.2 0 0 1-.173.18q-.035.029-.075.058a1.3 1.3 0 0 1-.251-.243 1 1 0 0 1-.076-.107c-.124-.193-.177-.363-.337-.444-.034-.016-.104-.008-.2.022q-.094.03-.216.087-.06.028-.125.063l-.13.074q-.067.04-.136.086a3 3 0 0 0-.135.096 3 3 0 0 0-.26.219 2 2 0 0 0-.12.121 2 2 0 0 0-.106.128l-.002.002a2 2 0 0 0-.09.132l-.001.001a1.2 1.2 0 0 0-.105.212q-.013.036-.024.073c-1.139-.875-1.563-2.317-1.563-3.203 0-.334.109-.457.305-.454m.836 10.354c.824-1.19.766-2.082-.365-3.194-1.13-1.112-1.789-2.738-1.789-2.738s-.246-.945-.806-.858-.97 1.499.202 2.362c1.173.864-.233 1.45-.685.64-.45-.812-1.683-2.896-2.322-3.295s-1.089-.175-.938.647 2.822 2.813 2.562 3.244-1.176-.506-1.176-.506-2.866-2.567-3.49-1.898.473 1.23 2.037 2.16c1.564.932 1.686 1.178 1.464 1.53s-3.675-2.511-4-1.297c-.323 1.214 3.524 1.567 3.287 2.405-.238.839-2.71-1.587-3.216-.642-.506.946 3.49 2.056 3.522 2.064 1.29.33 4.568 1.028 5.713-.624m5.349 0c-.824-1.19-.766-2.082.365-3.194 1.13-1.112 1.789-2.738 1.789-2.738s.246-.945.806-.858.97 1.499-.202 2.362c-1.173.864.233 1.45.685.64.451-.812 1.683-2.896 2.322-3.295s1.089-.175.938.647-2.822 2.813-2.562 3.244 1.176-.506 1.176-.506 2.866-2.567 3.49-1.898-.473 1.23-2.037 2.16c-1.564.932-1.686 1.178-1.464 1.53s3.675-2.511 4-1.297c.323 1.214-3.524 1.567-3.287 2.405.238.839 2.71-1.587 3.216-.642.506.946-3.49 2.056-3.522 2.064-1.29.33-4.568 1.028-5.713-.624"
+    />
+  </svg>
+{:else}
+  <svg class="w-6 h-6 {className}" viewBox="0 0 24 24" fill="currentColor">
+    <path
+      d="M12 2C6.48 2 2 6.48 2 12s4.48 10 10 10 10-4.48 10-10S17.52 2 12 2zm-2 15l-5-5 1.41-1.41L10 14.17l7.59-7.59L19 8l-9 9z"
+    />
+  </svg>
+{/if}

dashboard/src/lib/components/FamilySidebar.svelte

@@ -0,0 +1,142 @@
+<script lang="ts">
+  import FamilyLogos from "./FamilyLogos.svelte";
+
+  type FamilySidebarProps = {
+    families: string[];
+    selectedFamily: string | null;
+    hasFavorites: boolean;
+    onSelect: (family: string | null) => void;
+  };
+
+  let { families, selectedFamily, hasFavorites, onSelect }: FamilySidebarProps =
+    $props();
+
+  // Family display names
+  const familyNames: Record<string, string> = {
+    favorites: "Favorites",
+    huggingface: "Hub",
+    llama: "Meta",
+    qwen: "Qwen",
+    deepseek: "DeepSeek",
+    "gpt-oss": "OpenAI",
+    glm: "GLM",
+    minimax: "MiniMax",
+    kimi: "Kimi",
+  };
+
+  function getFamilyName(family: string): string {
+    return (
+      familyNames[family] || family.charAt(0).toUpperCase() + family.slice(1)
+    );
+  }
+</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-[64px]"
+>
+  <!-- All models (no filter) -->
+  <button
+    type="button"
+    onclick={() => onSelect(null)}
+    class="group flex flex-col items-center justify-center p-2 rounded transition-all duration-200 cursor-pointer {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
+        ? 'text-exo-yellow'
+        : 'text-white/40 group-hover:text-white/60'}">All</span
+    >
+  </button>
+
+  <!-- Favorites (only show if has favorites) -->
+  {#if hasFavorites}
+    <button
+      type="button"
+      onclick={() => onSelect("favorites")}
+      class="group flex flex-col items-center justify-center p-2 rounded transition-all duration-200 cursor-pointer {selectedFamily ===
+      'favorites'
+        ? 'bg-exo-yellow/20 border-l-2 border-exo-yellow'
+        : 'hover:bg-white/5 border-l-2 border-transparent'}"
+      title="Show favorited models"
+    >
+      <FamilyLogos
+        family="favorites"
+        class={selectedFamily === "favorites"
+          ? "text-amber-400"
+          : "text-white/50 group-hover:text-amber-400/70"}
+      />
+      <span
+        class="text-[9px] font-mono mt-0.5 {selectedFamily === 'favorites'
+          ? 'text-amber-400'
+          : 'text-white/40 group-hover:text-white/60'}">Faves</span
+      >
+    </button>
+  {/if}
+
+  <!-- HuggingFace Hub -->
+  <button
+    type="button"
+    onclick={() => onSelect("huggingface")}
+    class="group flex flex-col items-center justify-center p-2 rounded transition-all duration-200 cursor-pointer {selectedFamily ===
+    'huggingface'
+      ? 'bg-orange-500/20 border-l-2 border-orange-400'
+      : 'hover:bg-white/5 border-l-2 border-transparent'}"
+    title="Browse and add models from Hugging Face"
+  >
+    <FamilyLogos
+      family="huggingface"
+      class={selectedFamily === "huggingface"
+        ? "text-orange-400"
+        : "text-white/50 group-hover:text-orange-400/70"}
+    />
+    <span
+      class="text-[9px] font-mono mt-0.5 {selectedFamily === 'huggingface'
+        ? 'text-orange-400'
+        : 'text-white/40 group-hover:text-white/60'}">Hub</span
+    >
+  </button>
+
+  <div class="h-px bg-exo-yellow/10 my-1"></div>
+
+  <!-- Model families -->
+  {#each families as family}
+    <button
+      type="button"
+      onclick={() => onSelect(family)}
+      class="group flex flex-col items-center justify-center p-2 rounded transition-all duration-200 cursor-pointer {selectedFamily ===
+      family
+        ? 'bg-exo-yellow/20 border-l-2 border-exo-yellow'
+        : 'hover:bg-white/5 border-l-2 border-transparent'}"
+      title={getFamilyName(family)}
+    >
+      <FamilyLogos
+        {family}
+        class={selectedFamily === family
+          ? "text-exo-yellow"
+          : "text-white/50 group-hover:text-white/70"}
+      />
+      <span
+        class="text-[9px] font-mono mt-0.5 truncate max-w-full {selectedFamily ===
+        family
+          ? 'text-exo-yellow'
+          : 'text-white/40 group-hover:text-white/60'}"
+      >
+        {getFamilyName(family)}
+      </span>
+    </button>
+  {/each}
+</div>

dashboard/src/lib/components/HuggingFaceResultItem.svelte

@@ -0,0 +1,127 @@
+<script lang="ts">
+  interface HuggingFaceModel {
+    id: string;
+    author: string;
+    downloads: number;
+    likes: number;
+    last_modified: string;
+    tags: string[];
+  }
+
+  type HuggingFaceResultItemProps = {
+    model: HuggingFaceModel;
+    isAdded: boolean;
+    isAdding: boolean;
+    onAdd: () => void;
+    onSelect: () => void;
+  };
+
+  let {
+    model,
+    isAdded,
+    isAdding,
+    onAdd,
+    onSelect,
+  }: HuggingFaceResultItemProps = $props();
+
+  function formatNumber(num: number): string {
+    if (num >= 1000000) {
+      return `${(num / 1000000).toFixed(1)}M`;
+    } else if (num >= 1000) {
+      return `${(num / 1000).toFixed(1)}k`;
+    }
+    return num.toString();
+  }
+
+  // Extract model name from full ID (e.g., "mlx-community/Llama-3.2-1B" -> "Llama-3.2-1B")
+  const modelName = $derived(model.id.split("/").pop() || model.id);
+</script>
+
+<div
+  class="flex items-center justify-between gap-3 px-3 py-2.5 hover:bg-white/5 transition-colors border-b border-white/5 last:border-b-0"
+>
+  <div class="flex-1 min-w-0">
+    <div class="flex items-center gap-2">
+      <span class="text-sm font-mono text-white truncate" title={model.id}
+        >{modelName}</span
+      >
+      {#if isAdded}
+        <span
+          class="px-1.5 py-0.5 text-[10px] font-mono bg-green-500/20 text-green-400 rounded"
+          >Added</span
+        >
+      {/if}
+    </div>
+    <div class="flex items-center gap-3 mt-0.5 text-xs text-white/40">
+      <span class="truncate">{model.author}</span>
+      <span
+        class="flex items-center gap-1 shrink-0"
+        title="Downloads in the last 30 days"
+      >
+        <svg
+          class="w-3 h-3"
+          fill="none"
+          stroke="currentColor"
+          viewBox="0 0 24 24"
+        >
+          <path
+            stroke-linecap="round"
+            stroke-linejoin="round"
+            stroke-width="2"
+            d="M4 16v1a3 3 0 003 3h10a3 3 0 003-3v-1m-4-4l-4 4m0 0l-4-4m4 4V4"
+          />
+        </svg>
+        {formatNumber(model.downloads)}
+      </span>
+      <span
+        class="flex items-center gap-1 shrink-0"
+        title="Community likes on Hugging Face"
+      >
+        <svg
+          class="w-3 h-3"
+          fill="none"
+          stroke="currentColor"
+          viewBox="0 0 24 24"
+        >
+          <path
+            stroke-linecap="round"
+            stroke-linejoin="round"
+            stroke-width="2"
+            d="M4.318 6.318a4.5 4.5 0 000 6.364L12 20.364l7.682-7.682a4.5 4.5 0 00-6.364-6.364L12 7.636l-1.318-1.318a4.5 4.5 0 00-6.364 0z"
+          />
+        </svg>
+        {formatNumber(model.likes)}
+      </span>
+    </div>
+  </div>
+
+  <div class="flex items-center gap-2 shrink-0">
+    {#if isAdded}
+      <button
+        type="button"
+        onclick={onSelect}
+        class="px-3 py-1.5 text-xs font-mono tracking-wider uppercase bg-exo-yellow/10 text-exo-yellow border border-exo-yellow/30 hover:bg-exo-yellow/20 transition-colors rounded cursor-pointer"
+      >
+        Select
+      </button>
+    {:else}
+      <button
+        type="button"
+        onclick={onAdd}
+        disabled={isAdding}
+        class="px-3 py-1.5 text-xs font-mono tracking-wider uppercase bg-orange-500/10 text-orange-400 border border-orange-400/30 hover:bg-orange-500/20 transition-colors rounded cursor-pointer disabled:opacity-50 disabled:cursor-not-allowed"
+      >
+        {#if isAdding}
+          <span class="flex items-center gap-1.5">
+            <span
+              class="w-3 h-3 border-2 border-orange-400 border-t-transparent rounded-full animate-spin"
+            ></span>
+            Adding...
+          </span>
+        {:else}
+          + Add
+        {/if}
+      </button>
+    {/if}
+  </div>
+</div>

dashboard/src/lib/components/ModelFilterPopover.svelte

@@ -0,0 +1,182 @@
+<script lang="ts">
+  import { fly } from "svelte/transition";
+  import { cubicOut } from "svelte/easing";
+
+  interface FilterState {
+    capabilities: string[];
+    sizeRange: { min: number; max: number } | null;
+  }
+
+  type ModelFilterPopoverProps = {
+    filters: FilterState;
+    onChange: (filters: FilterState) => void;
+    onClear: () => void;
+    onClose: () => void;
+  };
+
+  let { filters, onChange, onClear, onClose }: ModelFilterPopoverProps =
+    $props();
+
+  // Available capabilities
+  const availableCapabilities = [
+    { id: "text", label: "Text" },
+    { id: "thinking", label: "Thinking" },
+    { id: "code", label: "Code" },
+    { id: "vision", label: "Vision" },
+  ];
+
+  // Size ranges
+  const sizeRanges = [
+    { label: "< 10GB", min: 0, max: 10 },
+    { label: "10-50GB", min: 10, max: 50 },
+    { label: "50-200GB", min: 50, max: 200 },
+    { label: "> 200GB", min: 200, max: 10000 },
+  ];
+
+  function toggleCapability(cap: string) {
+    const next = filters.capabilities.includes(cap)
+      ? filters.capabilities.filter((c) => c !== cap)
+      : [...filters.capabilities, cap];
+    onChange({ ...filters, capabilities: next });
+  }
+
+  function selectSizeRange(range: { min: number; max: number } | null) {
+    // Toggle off if same range is clicked
+    if (
+      filters.sizeRange &&
+      range &&
+      filters.sizeRange.min === range.min &&
+      filters.sizeRange.max === range.max
+    ) {
+      onChange({ ...filters, sizeRange: null });
+    } else {
+      onChange({ ...filters, sizeRange: range });
+    }
+  }
+
+  function handleClickOutside(e: MouseEvent) {
+    const target = e.target as HTMLElement;
+    if (
+      !target.closest(".filter-popover") &&
+      !target.closest(".filter-toggle")
+    ) {
+      onClose();
+    }
+  }
+</script>
+
+<svelte:window onclick={handleClickOutside} />
+
+<!-- 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"
+  transition:fly={{ y: -10, duration: 200, easing: cubicOut }}
+  onclick={(e) => e.stopPropagation()}
+  role="dialog"
+  aria-label="Filter options"
+>
+  <div class="p-3 space-y-4">
+    <!-- Capabilities -->
+    <div>
+      <h4 class="text-xs font-mono text-white/50 mb-2">Capabilities</h4>
+      <div class="flex flex-wrap gap-1.5">
+        {#each availableCapabilities as cap}
+          {@const isSelected = filters.capabilities.includes(cap.id)}
+          <button
+            type="button"
+            class="px-2 py-1 text-xs font-mono rounded transition-colors {isSelected
+              ? 'bg-exo-yellow/20 text-exo-yellow border border-exo-yellow/30'
+              : 'bg-white/5 text-white/60 hover:bg-white/10 border border-transparent'}"
+            onclick={() => toggleCapability(cap.id)}
+          >
+            {#if cap.id === "text"}
+              <svg
+                class="w-3.5 h-3.5 inline-block"
+                viewBox="0 0 24 24"
+                fill="none"
+                stroke="currentColor"
+                stroke-width="1.5"
+                ><path
+                  d="M21 15a2 2 0 0 1-2 2H7l-4 4V5a2 2 0 0 1 2-2h14a2 2 0 0 1 2 2z"
+                  stroke-linecap="round"
+                  stroke-linejoin="round"
+                /></svg
+              >
+            {:else if cap.id === "thinking"}
+              <svg
+                class="w-3.5 h-3.5 inline-block"
+                viewBox="0 0 24 24"
+                fill="none"
+                stroke="currentColor"
+                stroke-width="1.5"
+                ><path
+                  d="M12 2a7 7 0 0 0-7 7c0 2.38 1.19 4.47 3 5.74V17a1 1 0 0 0 1 1h6a1 1 0 0 0 1-1v-2.26c1.81-1.27 3-3.36 3-5.74a7 7 0 0 0-7-7zM9 20h6M10 22h4"
+                  stroke-linecap="round"
+                  stroke-linejoin="round"
+                /></svg
+              >
+            {:else if cap.id === "code"}
+              <svg
+                class="w-3.5 h-3.5 inline-block"
+                viewBox="0 0 24 24"
+                fill="none"
+                stroke="currentColor"
+                stroke-width="1.5"
+                ><path
+                  d="M16 18l6-6-6-6M8 6l-6 6 6 6"
+                  stroke-linecap="round"
+                  stroke-linejoin="round"
+                /></svg
+              >
+            {:else if cap.id === "vision"}
+              <svg
+                class="w-3.5 h-3.5 inline-block"
+                viewBox="0 0 24 24"
+                fill="none"
+                stroke="currentColor"
+                stroke-width="1.5"
+                ><path
+                  d="M1 12s4-8 11-8 11 8 11 8-4 8-11 8-11-8-11-8z"
+                  stroke-linecap="round"
+                  stroke-linejoin="round"
+                /><circle cx="12" cy="12" r="3" /></svg
+              >
+            {/if}
+            <span class="ml-1">{cap.label}</span>
+          </button>
+        {/each}
+      </div>
+    </div>
+
+    <!-- Size range -->
+    <div>
+      <h4 class="text-xs font-mono text-white/50 mb-2">Model Size</h4>
+      <div class="flex flex-wrap gap-1.5">
+        {#each sizeRanges as range}
+          {@const isSelected =
+            filters.sizeRange &&
+            filters.sizeRange.min === range.min &&
+            filters.sizeRange.max === range.max}
+          <button
+            type="button"
+            class="px-2 py-1 text-xs font-mono rounded transition-colors {isSelected
+              ? 'bg-exo-yellow/20 text-exo-yellow border border-exo-yellow/30'
+              : 'bg-white/5 text-white/60 hover:bg-white/10 border border-transparent'}"
+            onclick={() => selectSizeRange(range)}
+          >
+            {range.label}
+          </button>
+        {/each}
+      </div>
+    </div>
+
+    <!-- Clear button -->
+    <button
+      type="button"
+      class="w-full py-1.5 text-xs font-mono text-white/50 hover:text-white/70 hover:bg-white/5 rounded transition-colors"
+      onclick={onClear}
+    >
+      Clear all filters
+    </button>
+  </div>
+</div>

dashboard/src/lib/components/ModelPickerGroup.svelte

@@ -0,0 +1,324 @@
+<script lang="ts">
+  interface ModelInfo {
+    id: string;
+    name?: string;
+    storage_size_megabytes?: number;
+    base_model?: string;
+    quantization?: string;
+    supports_tensor?: boolean;
+    capabilities?: string[];
+    family?: string;
+    is_custom?: boolean;
+  }
+
+  interface ModelGroup {
+    id: string;
+    name: string;
+    capabilities: string[];
+    family: string;
+    variants: ModelInfo[];
+    smallestVariant: ModelInfo;
+    hasMultipleVariants: boolean;
+  }
+
+  type ModelPickerGroupProps = {
+    group: ModelGroup;
+    isExpanded: boolean;
+    isFavorite: boolean;
+    selectedModelId: string | null;
+    canModelFit: (id: string) => boolean;
+    onToggleExpand: () => void;
+    onSelectModel: (modelId: string) => void;
+    onToggleFavorite: (baseModelId: string) => void;
+    onShowInfo: (group: ModelGroup) => void;
+  };
+
+  let {
+    group,
+    isExpanded,
+    isFavorite,
+    selectedModelId,
+    canModelFit,
+    onToggleExpand,
+    onSelectModel,
+    onToggleFavorite,
+    onShowInfo,
+  }: ModelPickerGroupProps = $props();
+
+  // Format storage size
+  function formatSize(mb: number | undefined): string {
+    if (!mb) return "";
+    if (mb >= 1024) {
+      return `${(mb / 1024).toFixed(0)}GB`;
+    }
+    return `${mb}MB`;
+  }
+
+  // Check if any variant can fit
+  const anyVariantFits = $derived(
+    group.variants.some((v) => canModelFit(v.id)),
+  );
+
+  // Check if this group's model is currently selected (for single-variant groups)
+  const isMainSelected = $derived(
+    !group.hasMultipleVariants &&
+      group.variants.some((v) => v.id === selectedModelId),
+  );
+</script>
+
+<div
+  class="border-b border-white/5 last:border-b-0 {!anyVariantFits
+    ? 'opacity-50'
+    : ''}"
+>
+  <!-- Main row -->
+  <div
+    class="flex items-center gap-2 px-3 py-2.5 transition-colors {anyVariantFits
+      ? 'hover:bg-white/5 cursor-pointer'
+      : 'cursor-not-allowed'} {isMainSelected
+      ? 'bg-exo-yellow/10 border-l-2 border-exo-yellow'
+      : 'border-l-2 border-transparent'}"
+    onclick={() => {
+      if (group.hasMultipleVariants) {
+        onToggleExpand();
+      } else {
+        const modelId = group.variants[0]?.id;
+        if (modelId && canModelFit(modelId)) {
+          onSelectModel(modelId);
+        }
+      }
+    }}
+    role="button"
+    tabindex="0"
+    onkeydown={(e) => {
+      if (e.key === "Enter" || e.key === " ") {
+        e.preventDefault();
+        if (group.hasMultipleVariants) {
+          onToggleExpand();
+        } else {
+          const modelId = group.variants[0]?.id;
+          if (modelId && canModelFit(modelId)) {
+            onSelectModel(modelId);
+          }
+        }
+      }
+    }}
+  >
+    <!-- Expand/collapse chevron (for groups with variants) -->
+    {#if group.hasMultipleVariants}
+      <svg
+        class="w-4 h-4 text-white/40 transition-transform duration-200 flex-shrink-0 {isExpanded
+          ? 'rotate-90'
+          : ''}"
+        viewBox="0 0 24 24"
+        fill="currentColor"
+      >
+        <path d="M8.59 16.59L13.17 12 8.59 7.41 10 6l6 6-6 6-1.41-1.41z" />
+      </svg>
+    {:else}
+      <div class="w-4 flex-shrink-0"></div>
+    {/if}
+
+    <!-- Model name -->
+    <div class="flex-1 min-w-0">
+      <div class="flex items-center gap-2">
+        <span class="font-mono text-sm text-white truncate">
+          {group.name}
+        </span>
+        <!-- Capability icons -->
+        {#each group.capabilities.filter((c) => c !== "text") as cap}
+          {#if cap === "thinking"}
+            <svg
+              class="w-3.5 h-3.5 text-white/40 flex-shrink-0"
+              viewBox="0 0 24 24"
+              fill="none"
+              stroke="currentColor"
+              stroke-width="1.5"
+              title="Supports Thinking"
+            >
+              <path
+                d="M12 2a7 7 0 0 0-7 7c0 2.38 1.19 4.47 3 5.74V17a1 1 0 0 0 1 1h6a1 1 0 0 0 1-1v-2.26c1.81-1.27 3-3.36 3-5.74a7 7 0 0 0-7-7zM9 20h6M10 22h4"
+                stroke-linecap="round"
+                stroke-linejoin="round"
+              />
+            </svg>
+          {:else if cap === "code"}
+            <svg
+              class="w-3.5 h-3.5 text-white/40 flex-shrink-0"
+              viewBox="0 0 24 24"
+              fill="none"
+              stroke="currentColor"
+              stroke-width="1.5"
+              title="Supports code generation"
+            >
+              <path
+                d="M16 18l6-6-6-6M8 6l-6 6 6 6"
+                stroke-linecap="round"
+                stroke-linejoin="round"
+              />
+            </svg>
+          {:else if cap === "vision"}
+            <svg
+              class="w-3.5 h-3.5 text-white/40 flex-shrink-0"
+              viewBox="0 0 24 24"
+              fill="none"
+              stroke="currentColor"
+              stroke-width="1.5"
+              title="Supports image input"
+            >
+              <path
+                d="M1 12s4-8 11-8 11 8 11 8-4 8-11 8-11-8-11-8z"
+                stroke-linecap="round"
+                stroke-linejoin="round"
+              />
+              <circle cx="12" cy="12" r="3" />
+            </svg>
+          {:else if cap === "image_gen"}
+            <svg
+              class="w-3.5 h-3.5 text-white/40 flex-shrink-0"
+              viewBox="0 0 24 24"
+              fill="none"
+              stroke="currentColor"
+              stroke-width="1.5"
+              title="Supports image generation"
+            >
+              <rect x="3" y="3" width="18" height="18" rx="2" ry="2" />
+              <circle cx="8.5" cy="8.5" r="1.5" />
+              <path d="M21 15l-5-5L5 21" />
+            </svg>
+          {/if}
+        {/each}
+      </div>
+    </div>
+
+    <!-- Size indicator (smallest variant) -->
+    {#if !group.hasMultipleVariants && group.smallestVariant?.storage_size_megabytes}
+      <span class="text-xs font-mono text-white/30 flex-shrink-0">
+        {formatSize(group.smallestVariant.storage_size_megabytes)}
+      </span>
+    {/if}
+
+    <!-- Variant count -->
+    {#if group.hasMultipleVariants}
+      <span class="text-xs font-mono text-white/30 flex-shrink-0">
+        {group.variants.length} variants
+      </span>
+    {/if}
+
+    <!-- Check mark if selected (single-variant) -->
+    {#if isMainSelected}
+      <svg
+        class="w-4 h-4 text-exo-yellow flex-shrink-0"
+        viewBox="0 0 24 24"
+        fill="currentColor"
+      >
+        <path d="M9 16.17L4.83 12l-1.42 1.41L9 19 21 7l-1.41-1.41L9 16.17z" />
+      </svg>
+    {/if}
+
+    <!-- Favorite star -->
+    <button
+      type="button"
+      class="p-1 rounded hover:bg-white/10 transition-colors flex-shrink-0"
+      onclick={(e) => {
+        e.stopPropagation();
+        onToggleFavorite(group.id);
+      }}
+      title={isFavorite ? "Remove from favorites" : "Add to favorites"}
+    >
+      {#if isFavorite}
+        <svg
+          class="w-4 h-4 text-amber-400"
+          viewBox="0 0 24 24"
+          fill="currentColor"
+        >
+          <path
+            d="M12 2l3.09 6.26L22 9.27l-5 4.87 1.18 6.88L12 17.77l-6.18 3.25L7 14.14 2 9.27l6.91-1.01L12 2z"
+          />
+        </svg>
+      {:else}
+        <svg
+          class="w-4 h-4 text-white/30 hover:text-white/50"
+          viewBox="0 0 24 24"
+          fill="none"
+          stroke="currentColor"
+          stroke-width="2"
+        >
+          <path
+            d="M12 2l3.09 6.26L22 9.27l-5 4.87 1.18 6.88L12 17.77l-6.18 3.25L7 14.14 2 9.27l6.91-1.01L12 2z"
+          />
+        </svg>
+      {/if}
+    </button>
+
+    <!-- Info button -->
+    <button
+      type="button"
+      class="p-1 rounded hover:bg-white/10 transition-colors flex-shrink-0"
+      onclick={(e) => {
+        e.stopPropagation();
+        onShowInfo(group);
+      }}
+      title="View model details"
+    >
+      <svg
+        class="w-4 h-4 text-white/30 hover:text-white/50"
+        viewBox="0 0 24 24"
+        fill="currentColor"
+      >
+        <path
+          d="M12 2C6.48 2 2 6.48 2 12s4.48 10 10 10 10-4.48 10-10S17.52 2 12 2zm1 15h-2v-6h2v6zm0-8h-2V7h2v2z"
+        />
+      </svg>
+    </button>
+  </div>
+
+  <!-- Expanded variants -->
+  {#if isExpanded && group.hasMultipleVariants}
+    <div class="bg-black/20 border-t border-white/5">
+      {#each group.variants as variant}
+        {@const modelCanFit = canModelFit(variant.id)}
+        {@const isSelected = selectedModelId === variant.id}
+        <button
+          type="button"
+          class="w-full flex items-center gap-3 px-3 py-2 pl-10 hover:bg-white/5 transition-colors text-left {!modelCanFit
+            ? 'opacity-50 cursor-not-allowed'
+            : 'cursor-pointer'} {isSelected
+            ? 'bg-exo-yellow/10 border-l-2 border-exo-yellow'
+            : 'border-l-2 border-transparent'}"
+          disabled={!modelCanFit}
+          onclick={() => {
+            if (modelCanFit) {
+              onSelectModel(variant.id);
+            }
+          }}
+        >
+          <!-- Quantization badge -->
+          <span
+            class="text-xs font-mono px-1.5 py-0.5 rounded bg-white/10 text-white/70 flex-shrink-0"
+          >
+            {variant.quantization || "default"}
+          </span>
+
+          <!-- Size -->
+          <span class="text-xs font-mono text-white/40 flex-1">
+            {formatSize(variant.storage_size_megabytes)}
+          </span>
+
+          <!-- Check mark if selected -->
+          {#if isSelected}
+            <svg
+              class="w-4 h-4 text-exo-yellow"
+              viewBox="0 0 24 24"
+              fill="currentColor"
+            >
+              <path
+                d="M9 16.17L4.83 12l-1.42 1.41L9 19 21 7l-1.41-1.41L9 16.17z"
+              />
+            </svg>
+          {/if}
+        </button>
+      {/each}
+    </div>
+  {/if}
+</div>

dashboard/src/lib/components/ModelPickerModal.svelte

@@ -0,0 +1,748 @@
+<script lang="ts">
+  import { fade, fly } from "svelte/transition";
+  import { cubicOut } from "svelte/easing";
+  import FamilySidebar from "./FamilySidebar.svelte";
+  import ModelPickerGroup from "./ModelPickerGroup.svelte";
+  import ModelFilterPopover from "./ModelFilterPopover.svelte";
+  import HuggingFaceResultItem from "./HuggingFaceResultItem.svelte";
+
+  interface ModelInfo {
+    id: string;
+    name?: string;
+    storage_size_megabytes?: number;
+    base_model?: string;
+    quantization?: string;
+    supports_tensor?: boolean;
+    capabilities?: string[];
+    family?: string;
+    is_custom?: boolean;
+    tasks?: string[];
+    hugging_face_id?: string;
+  }
+
+  interface ModelGroup {
+    id: string;
+    name: string;
+    capabilities: string[];
+    family: string;
+    variants: ModelInfo[];
+    smallestVariant: ModelInfo;
+    hasMultipleVariants: boolean;
+  }
+
+  interface FilterState {
+    capabilities: string[];
+    sizeRange: { min: number; max: number } | null;
+  }
+
+  interface HuggingFaceModel {
+    id: string;
+    author: string;
+    downloads: number;
+    likes: number;
+    last_modified: string;
+    tags: string[];
+  }
+
+  type ModelPickerModalProps = {
+    isOpen: boolean;
+    models: ModelInfo[];
+    selectedModelId: string | null;
+    favorites: Set<string>;
+    existingModelIds: Set<string>;
+    canModelFit: (modelId: string) => boolean;
+    onSelect: (modelId: string) => void;
+    onClose: () => void;
+    onToggleFavorite: (baseModelId: string) => void;
+    onAddModel: (modelId: string) => Promise<void>;
+    onDeleteModel: (modelId: string) => Promise<void>;
+    totalMemoryGB: number;
+    usedMemoryGB: number;
+  };
+
+  let {
+    isOpen,
+    models,
+    selectedModelId,
+    favorites,
+    existingModelIds,
+    canModelFit,
+    onSelect,
+    onClose,
+    onToggleFavorite,
+    onAddModel,
+    onDeleteModel,
+    totalMemoryGB,
+    usedMemoryGB,
+  }: ModelPickerModalProps = $props();
+
+  // Local state
+  let searchQuery = $state("");
+  let selectedFamily = $state<string | null>(null);
+  let expandedGroups = $state<Set<string>>(new Set());
+  let showFilters = $state(false);
+  let filters = $state<FilterState>({ capabilities: [], sizeRange: null });
+  let infoGroup = $state<ModelGroup | null>(null);
+
+  // HuggingFace Hub state
+  let hfSearchQuery = $state("");
+  let hfSearchResults = $state<HuggingFaceModel[]>([]);
+  let hfTrendingModels = $state<HuggingFaceModel[]>([]);
+  let hfIsSearching = $state(false);
+  let hfIsLoadingTrending = $state(false);
+  let addingModelId = $state<string | null>(null);
+  let hfSearchDebounceTimer: ReturnType<typeof setTimeout> | null = null;
+  let manualModelId = $state("");
+  let addModelError = $state<string | null>(null);
+
+  // Reset state when modal opens
+  $effect(() => {
+    if (isOpen) {
+      searchQuery = "";
+      selectedFamily = null;
+      expandedGroups = new Set();
+      showFilters = false;
+      hfSearchQuery = "";
+      hfSearchResults = [];
+      manualModelId = "";
+      addModelError = null;
+    }
+  });
+
+  // Fetch trending models when HuggingFace is selected
+  $effect(() => {
+    if (
+      selectedFamily === "huggingface" &&
+      hfTrendingModels.length === 0 &&
+      !hfIsLoadingTrending
+    ) {
+      fetchTrendingModels();
+    }
+  });
+
+  async function fetchTrendingModels() {
+    hfIsLoadingTrending = true;
+    try {
+      const response = await fetch("/models/search?query=&limit=20");
+      if (response.ok) {
+        hfTrendingModels = await response.json();
+      }
+    } catch (error) {
+      console.error("Failed to fetch trending models:", error);
+    } finally {
+      hfIsLoadingTrending = false;
+    }
+  }
+
+  async function searchHuggingFace(query: string) {
+    if (query.length < 2) {
+      hfSearchResults = [];
+      return;
+    }
+
+    hfIsSearching = true;
+    try {
+      const response = await fetch(
+        `/models/search?query=${encodeURIComponent(query)}&limit=20`,
+      );
+      if (response.ok) {
+        hfSearchResults = await response.json();
+      } else {
+        hfSearchResults = [];
+      }
+    } catch (error) {
+      console.error("Failed to search models:", error);
+      hfSearchResults = [];
+    } finally {
+      hfIsSearching = false;
+    }
+  }
+
+  function handleHfSearchInput(query: string) {
+    hfSearchQuery = query;
+    addModelError = null;
+
+    if (hfSearchDebounceTimer) {
+      clearTimeout(hfSearchDebounceTimer);
+    }
+
+    if (query.length >= 2) {
+      hfSearchDebounceTimer = setTimeout(() => {
+        searchHuggingFace(query);
+      }, 300);
+    } else {
+      hfSearchResults = [];
+    }
+  }
+
+  async function handleAddModel(modelId: string) {
+    addingModelId = modelId;
+    addModelError = null;
+    try {
+      await onAddModel(modelId);
+    } catch (error) {
+      addModelError =
+        error instanceof Error ? error.message : "Failed to add model";
+    } finally {
+      addingModelId = null;
+    }
+  }
+
+  async function handleAddManualModel() {
+    if (!manualModelId.trim()) return;
+    await handleAddModel(manualModelId.trim());
+    if (!addModelError) {
+      manualModelId = "";
+    }
+  }
+
+  function handleSelectHfModel(modelId: string) {
+    onSelect(modelId);
+    onClose();
+  }
+
+  // Models to display in HuggingFace view
+  const hfDisplayModels = $derived.by((): HuggingFaceModel[] => {
+    if (hfSearchQuery.length >= 2) {
+      return hfSearchResults;
+    }
+    return hfTrendingModels;
+  });
+
+  // Group models by base_model
+  const groupedModels = $derived.by((): ModelGroup[] => {
+    const groups = new Map<string, ModelGroup>();
+
+    for (const model of models) {
+      const groupId = model.base_model || model.id;
+      const groupName = model.base_model || model.name || model.id;
+
+      if (!groups.has(groupId)) {
+        groups.set(groupId, {
+          id: groupId,
+          name: groupName,
+          capabilities: model.capabilities || ["text"],
+          family: model.family || "",
+          variants: [],
+          smallestVariant: model,
+          hasMultipleVariants: false,
+        });
+      }
+
+      const group = groups.get(groupId)!;
+      group.variants.push(model);
+
+      // Track smallest variant
+      if (
+        (model.storage_size_megabytes || 0) <
+        (group.smallestVariant.storage_size_megabytes || Infinity)
+      ) {
+        group.smallestVariant = model;
+      }
+
+      // Update capabilities if not set
+      if (
+        group.capabilities.length <= 1 &&
+        model.capabilities &&
+        model.capabilities.length > 1
+      ) {
+        group.capabilities = model.capabilities;
+      }
+      if (!group.family && model.family) {
+        group.family = model.family;
+      }
+    }
+
+    // Sort variants within each group by size
+    for (const group of groups.values()) {
+      group.variants.sort(
+        (a, b) =>
+          (a.storage_size_megabytes || 0) - (b.storage_size_megabytes || 0),
+      );
+      group.hasMultipleVariants = group.variants.length > 1;
+    }
+
+    // Convert to array and sort by smallest variant size (biggest first)
+    return Array.from(groups.values()).sort((a, b) => {
+      return (
+        (b.smallestVariant.storage_size_megabytes || 0) -
+        (a.smallestVariant.storage_size_megabytes || 0)
+      );
+    });
+  });
+
+  // Get unique families
+  const uniqueFamilies = $derived.by((): string[] => {
+    const families = new Set<string>();
+    for (const group of groupedModels) {
+      if (group.family) {
+        families.add(group.family);
+      }
+    }
+    const familyOrder = [
+      "kimi",
+      "qwen",
+      "glm",
+      "minimax",
+      "deepseek",
+      "gpt-oss",
+      "llama",
+    ];
+    return Array.from(families).sort((a, b) => {
+      const aIdx = familyOrder.indexOf(a);
+      const bIdx = familyOrder.indexOf(b);
+      if (aIdx === -1 && bIdx === -1) return a.localeCompare(b);
+      if (aIdx === -1) return 1;
+      if (bIdx === -1) return -1;
+      return aIdx - bIdx;
+    });
+  });
+
+  // Filter models based on search, family, and filters
+  const filteredGroups = $derived.by((): ModelGroup[] => {
+    let result: ModelGroup[] = [...groupedModels];
+
+    // Filter by family
+    if (selectedFamily === "favorites") {
+      result = result.filter((g) => favorites.has(g.id));
+    } else if (selectedFamily && selectedFamily !== "huggingface") {
+      result = result.filter((g) => g.family === selectedFamily);
+    }
+
+    // Filter by search query
+    if (searchQuery.trim()) {
+      const query = searchQuery.toLowerCase().trim();
+      result = result.filter(
+        (g) =>
+          g.name.toLowerCase().includes(query) ||
+          g.variants.some(
+            (v) =>
+              v.id.toLowerCase().includes(query) ||
+              (v.name || "").toLowerCase().includes(query),
+          ),
+      );
+    }
+
+    // Filter by capabilities
+    if (filters.capabilities.length > 0) {
+      result = result.filter((g) =>
+        filters.capabilities.every((cap) => g.capabilities.includes(cap)),
+      );
+    }
+
+    // Filter by size range
+    if (filters.sizeRange) {
+      const { min, max } = filters.sizeRange;
+      result = result.filter((g) => {
+        const sizeGB = (g.smallestVariant.storage_size_megabytes || 0) / 1024;
+        return sizeGB >= min && sizeGB <= max;
+      });
+    }
+
+    // Sort: models that fit first, then by size (largest first)
+    result.sort((a, b) => {
+      const aFits = a.variants.some((v) => canModelFit(v.id));
+      const bFits = b.variants.some((v) => canModelFit(v.id));
+
+      if (aFits && !bFits) return -1;
+      if (!aFits && bFits) return 1;
+
+      return (
+        (b.smallestVariant.storage_size_megabytes || 0) -
+        (a.smallestVariant.storage_size_megabytes || 0)
+      );
+    });
+
+    return result;
+  });
+
+  // Check if any favorites exist
+  const hasFavorites = $derived(favorites.size > 0);
+
+  function toggleGroupExpanded(groupId: string) {
+    const next = new Set(expandedGroups);
+    if (next.has(groupId)) {
+      next.delete(groupId);
+    } else {
+      next.add(groupId);
+    }
+    expandedGroups = next;
+  }
+
+  function handleSelect(modelId: string) {
+    onSelect(modelId);
+    onClose();
+  }
+
+  function handleKeydown(e: KeyboardEvent) {
+    if (e.key === "Escape") {
+      onClose();
+    }
+  }
+
+  function handleFiltersChange(newFilters: FilterState) {
+    filters = newFilters;
+  }
+
+  function clearFilters() {
+    filters = { capabilities: [], sizeRange: null };
+  }
+
+  const hasActiveFilters = $derived(
+    filters.capabilities.length > 0 || filters.sizeRange !== null,
+  );
+</script>
+
+<svelte:window onkeydown={handleKeydown} />
+
+{#if isOpen}
+  <!-- Backdrop -->
+  <div
+    class="fixed inset-0 z-50 bg-black/80 backdrop-blur-sm"
+    transition:fade={{ duration: 200 }}
+    onclick={onClose}
+    role="presentation"
+  ></div>
+
+  <!-- Modal -->
+  <div
+    class="fixed z-50 top-1/2 left-1/2 -translate-x-1/2 -translate-y-1/2 w-[min(90vw,600px)] h-[min(80vh,700px)] bg-exo-dark-gray border border-exo-yellow/10 rounded-lg shadow-2xl overflow-hidden flex flex-col"
+    transition:fly={{ y: 20, duration: 300, easing: cubicOut }}
+    role="dialog"
+    aria-modal="true"
+    aria-label="Select a model"
+  >
+    <!-- Header with search -->
+    <div
+      class="flex items-center gap-2 p-3 border-b border-exo-yellow/10 bg-exo-medium-gray/30"
+    >
+      {#if selectedFamily === "huggingface"}
+        <!-- HuggingFace search -->
+        <svg
+          class="w-5 h-5 text-orange-400/60 flex-shrink-0"
+          viewBox="0 0 24 24"
+          fill="none"
+          stroke="currentColor"
+          stroke-width="2"
+        >
+          <circle cx="11" cy="11" r="8" />
+          <path d="M21 21l-4.35-4.35" />
+        </svg>
+        <input
+          type="search"
+          class="flex-1 bg-transparent border-none outline-none text-sm font-mono text-white placeholder-white/40"
+          placeholder="Search mlx-community models..."
+          value={hfSearchQuery}
+          oninput={(e) => handleHfSearchInput(e.currentTarget.value)}
+        />
+        {#if hfIsSearching}
+          <div class="flex-shrink-0">
+            <span
+              class="w-4 h-4 border-2 border-orange-400 border-t-transparent rounded-full animate-spin block"
+            ></span>
+          </div>
+        {/if}
+      {:else}
+        <!-- Normal model search -->
+        <svg
+          class="w-5 h-5 text-white/40 flex-shrink-0"
+          viewBox="0 0 24 24"
+          fill="none"
+          stroke="currentColor"
+          stroke-width="2"
+        >
+          <circle cx="11" cy="11" r="8" />
+          <path d="M21 21l-4.35-4.35" />
+        </svg>
+        <input
+          type="search"
+          class="flex-1 bg-transparent border-none outline-none text-sm font-mono text-white placeholder-white/40"
+          placeholder="Search models..."
+          bind:value={searchQuery}
+        />
+        <!-- Cluster memory -->
+        <span
+          class="text-xs font-mono flex-shrink-0"
+          title="Cluster memory usage"
+          ><span class="text-exo-yellow">{Math.round(usedMemoryGB)}GB</span
+          ><span class="text-white/40">/{Math.round(totalMemoryGB)}GB</span
+          ></span
+        >
+        <!-- Filter button -->
+        <div class="relative filter-toggle">
+          <button
+            type="button"
+            class="p-1.5 rounded hover:bg-white/10 transition-colors {hasActiveFilters
+              ? 'text-exo-yellow'
+              : 'text-white/50'}"
+            onclick={() => (showFilters = !showFilters)}
+            title="Filter by capability or size"
+          >
+            <svg class="w-5 h-5" viewBox="0 0 24 24" fill="currentColor">
+              <path d="M10 18h4v-2h-4v2zM3 6v2h18V6H3zm3 7h12v-2H6v2z" />
+            </svg>
+          </button>
+          {#if showFilters}
+            <ModelFilterPopover
+              {filters}
+              onChange={handleFiltersChange}
+              onClear={clearFilters}
+              onClose={() => (showFilters = false)}
+            />
+          {/if}
+        </div>
+      {/if}
+      <!-- Close button -->
+      <button
+        type="button"
+        class="p-1.5 rounded hover:bg-white/10 transition-colors text-white/50 hover:text-white/70"
+        onclick={onClose}
+        title="Close model picker"
+      >
+        <svg class="w-5 h-5" viewBox="0 0 24 24" fill="currentColor">
+          <path
+            d="M19 6.41L17.59 5 12 10.59 6.41 5 5 6.41 10.59 12 5 17.59 6.41 19 12 13.41 17.59 19 19 17.59 13.41 12 19 6.41z"
+          />
+        </svg>
+      </button>
+    </div>
+
+    <!-- Body -->
+    <div class="flex flex-1 overflow-hidden">
+      <!-- Family sidebar -->
+      <FamilySidebar
+        families={uniqueFamilies}
+        {selectedFamily}
+        {hasFavorites}
+        onSelect={(family) => (selectedFamily = family)}
+      />
+
+      <!-- Model list -->
+      <div class="flex-1 overflow-y-auto flex flex-col">
+        {#if selectedFamily === "huggingface"}
+          <!-- HuggingFace Hub view -->
+          <div class="flex-1 flex flex-col min-h-0">
+            <!-- Section header -->
+            <div
+              class="sticky top-0 z-10 px-3 py-2 bg-exo-dark-gray/95 border-b border-exo-yellow/10"
+            >
+              <span class="text-xs font-mono text-white/40">
+                {#if hfSearchQuery.length >= 2}
+                  Search results for "{hfSearchQuery}"
+                {:else}
+                  Trending on mlx-community
+                {/if}
+              </span>
+            </div>
+
+            <!-- Results list -->
+            <div class="flex-1 overflow-y-auto">
+              {#if hfIsLoadingTrending && hfTrendingModels.length === 0}
+                <div
+                  class="flex items-center justify-center py-12 text-white/40"
+                >
+                  <span
+                    class="w-5 h-5 border-2 border-orange-400 border-t-transparent rounded-full animate-spin mr-2"
+                  ></span>
+                  <span class="font-mono text-sm"
+                    >Loading trending models...</span
+                  >
+                </div>
+              {:else if hfDisplayModels.length === 0}
+                <div
+                  class="flex flex-col items-center justify-center py-12 text-white/40"
+                >
+                  <svg
+                    class="w-10 h-10 mb-2"
+                    viewBox="0 0 24 24"
+                    fill="currentColor"
+                  >
+                    <path
+                      d="M12 2C6.48 2 2 6.48 2 12s4.48 10 10 10 10-4.48 10-10S17.52 2 12 2zm-2 13.5c-.83 0-1.5-.67-1.5-1.5s.67-1.5 1.5-1.5 1.5.67 1.5 1.5-.67 1.5-1.5 1.5zm4 0c-.83 0-1.5-.67-1.5-1.5s.67-1.5 1.5-1.5 1.5.67 1.5 1.5-.67 1.5-1.5 1.5zm2-4.5H8c0-2.21 1.79-4 4-4s4 1.79 4 4z"
+                    />
+                  </svg>
+                  <p class="font-mono text-sm">No models found</p>
+                  {#if hfSearchQuery}
+                    <p class="font-mono text-xs mt-1">
+                      Try a different search term
+                    </p>
+                  {/if}
+                </div>
+              {:else}
+                {#each hfDisplayModels as model}
+                  <HuggingFaceResultItem
+                    {model}
+                    isAdded={existingModelIds.has(model.id)}
+                    isAdding={addingModelId === model.id}
+                    onAdd={() => handleAddModel(model.id)}
+                    onSelect={() => handleSelectHfModel(model.id)}
+                  />
+                {/each}
+              {/if}
+            </div>
+
+            <!-- Manual input footer -->
+            <div
+              class="sticky bottom-0 border-t border-exo-yellow/10 bg-exo-dark-gray p-3"
+            >
+              {#if addModelError}
+                <div
+                  class="bg-red-500/10 border border-red-500/30 rounded px-3 py-2 mb-2"
+                >
+                  <p class="text-red-400 text-xs font-mono break-words">
+                    {addModelError}
+                  </p>
+                </div>
+              {/if}
+              <div class="flex gap-2">
+                <input
+                  type="text"
+                  class="flex-1 bg-exo-black/60 border border-exo-yellow/30 rounded px-3 py-1.5 text-xs font-mono text-white placeholder-white/30 focus:outline-none focus:border-exo-yellow/50"
+                  placeholder="Or paste model ID directly..."
+                  bind:value={manualModelId}
+                  onkeydown={(e) => {
+                    if (e.key === "Enter") handleAddManualModel();
+                  }}
+                />
+                <button
+                  type="button"
+                  onclick={handleAddManualModel}
+                  disabled={!manualModelId.trim() || addingModelId !== null}
+                  class="px-3 py-1.5 text-xs font-mono tracking-wider uppercase bg-orange-500/10 text-orange-400 border border-orange-400/30 hover:bg-orange-500/20 transition-colors rounded disabled:opacity-50 disabled:cursor-not-allowed"
+                >
+                  Add
+                </button>
+              </div>
+            </div>
+          </div>
+        {:else if filteredGroups.length === 0}
+          <div
+            class="flex flex-col items-center justify-center h-full text-white/40 p-8"
+          >
+            <svg class="w-12 h-12 mb-3" viewBox="0 0 24 24" fill="currentColor">
+              <path
+                d="M12 2C6.48 2 2 6.48 2 12s4.48 10 10 10 10-4.48 10-10S17.52 2 12 2zm-2 15l-5-5 1.41-1.41L10 14.17l7.59-7.59L19 8l-9 9z"
+              />
+            </svg>
+            <p class="font-mono text-sm">No models found</p>
+            {#if hasActiveFilters || searchQuery}
+              <button
+                type="button"
+                class="mt-2 text-xs text-exo-yellow hover:underline"
+                onclick={() => {
+                  searchQuery = "";
+                  clearFilters();
+                }}
+              >
+                Clear filters
+              </button>
+            {/if}
+          </div>
+        {:else}
+          {#each filteredGroups as group}
+            <ModelPickerGroup
+              {group}
+              isExpanded={expandedGroups.has(group.id)}
+              isFavorite={favorites.has(group.id)}
+              {selectedModelId}
+              {canModelFit}
+              onToggleExpand={() => toggleGroupExpanded(group.id)}
+              onSelectModel={handleSelect}
+              {onToggleFavorite}
+              onShowInfo={(g) => (infoGroup = g)}
+            />
+          {/each}
+        {/if}
+      </div>
+    </div>
+
+    <!-- Footer with active filters indicator -->
+    {#if hasActiveFilters}
+      <div
+        class="flex items-center gap-2 px-3 py-2 border-t border-exo-yellow/10 bg-exo-medium-gray/20 text-xs font-mono text-white/50"
+      >
+        <span>Filters:</span>
+        {#each filters.capabilities as cap}
+          <span class="px-1.5 py-0.5 bg-exo-yellow/20 text-exo-yellow rounded"
+            >{cap}</span
+          >
+        {/each}
+        {#if filters.sizeRange}
+          <span class="px-1.5 py-0.5 bg-exo-yellow/20 text-exo-yellow rounded">
+            {filters.sizeRange.min}GB - {filters.sizeRange.max}GB
+          </span>
+        {/if}
+        <button
+          type="button"
+          class="ml-auto text-white/40 hover:text-white/60"
+          onclick={clearFilters}
+        >
+          Clear all
+        </button>
+      </div>
+    {/if}
+  </div>
+
+  <!-- Info modal -->
+  {#if infoGroup}
+    <div
+      class="fixed inset-0 z-[60] bg-black/60"
+      transition:fade={{ duration: 150 }}
+      onclick={() => (infoGroup = null)}
+      role="presentation"
+    ></div>
+    <div
+      class="fixed z-[60] top-1/2 left-1/2 -translate-x-1/2 -translate-y-1/2 w-[min(80vw,400px)] bg-exo-dark-gray border border-exo-yellow/10 rounded-lg shadow-2xl p-4"
+      transition:fly={{ y: 10, duration: 200, easing: cubicOut }}
+      role="dialog"
+      aria-modal="true"
+    >
+      <div class="flex items-start justify-between mb-3">
+        <h3 class="font-mono text-lg text-white">{infoGroup.name}</h3>
+        <button
+          type="button"
+          class="p-1 rounded hover:bg-white/10 transition-colors text-white/50"
+          onclick={() => (infoGroup = null)}
+          title="Close model details"
+          aria-label="Close info dialog"
+        >
+          <svg class="w-4 h-4" viewBox="0 0 24 24" fill="currentColor">
+            <path
+              d="M19 6.41L17.59 5 12 10.59 6.41 5 5 6.41 10.59 12 5 17.59 6.41 19 12 13.41 17.59 19 19 17.59 13.41 12 19 6.41z"
+            />
+          </svg>
+        </button>
+      </div>
+      <div class="space-y-2 text-xs font-mono">
+        <div class="flex items-center gap-2">
+          <span class="text-white/40">Family:</span>
+          <span class="text-white/70">{infoGroup.family || "Unknown"}</span>
+        </div>
+        <div class="flex items-center gap-2">
+          <span class="text-white/40">Capabilities:</span>
+          <span class="text-white/70">{infoGroup.capabilities.join(", ")}</span>
+        </div>
+        <div class="flex items-center gap-2">
+          <span class="text-white/40">Variants:</span>
+          <span class="text-white/70">{infoGroup.variants.length}</span>
+        </div>
+        {#if infoGroup.variants.length > 0}
+          <div class="mt-3 pt-3 border-t border-exo-yellow/10">
+            <span class="text-white/40">Available quantizations:</span>
+            <div class="flex flex-wrap gap-1 mt-1">
+              {#each infoGroup.variants as variant}
+                <span
+                  class="px-1.5 py-0.5 bg-white/10 text-white/60 rounded text-[10px]"
+                >
+                  {variant.quantization || "default"} ({Math.round(
+                    (variant.storage_size_megabytes || 0) / 1024,
+                  )}GB)
+                </span>
+              {/each}
+            </div>
+          </div>
+        {/if}
+      </div>
+    </div>
+  {/if}
+{/if}

dashboard/src/lib/components/index.ts

@@ -6,3 +6,9 @@ export { default as ChatSidebar } from "./ChatSidebar.svelte";
 export { default as ModelCard } from "./ModelCard.svelte";
 export { default as MarkdownContent } from "./MarkdownContent.svelte";
 export { default as ImageParamsPanel } from "./ImageParamsPanel.svelte";
+export { default as FamilyLogos } from "./FamilyLogos.svelte";
+export { default as FamilySidebar } from "./FamilySidebar.svelte";
+export { default as HuggingFaceResultItem } from "./HuggingFaceResultItem.svelte";
+export { default as ModelFilterPopover } from "./ModelFilterPopover.svelte";
+export { default as ModelPickerGroup } from "./ModelPickerGroup.svelte";
+export { default as ModelPickerModal } from "./ModelPickerModal.svelte";

dashboard/src/lib/stores/app.svelte.ts

@@ -178,6 +178,36 @@ interface ImageApiResponse {
   data: Array<{ b64_json?: string; url?: string }>;
 }
 
+// Trace API response types
+export interface TraceCategoryStats {
+  totalUs: number;
+  count: number;
+  minUs: number;
+  maxUs: number;
+  avgUs: number;
+}
+
+export interface TraceRankStats {
+  byCategory: Record<string, TraceCategoryStats>;
+}
+
+export interface TraceStatsResponse {
+  taskId: string;
+  totalWallTimeUs: number;
+  byCategory: Record<string, TraceCategoryStats>;
+  byRank: Record<number, TraceRankStats>;
+}
+
+export interface TraceListItem {
+  taskId: string;
+  createdAt: string;
+  fileSize: number;
+}
+
+export interface TraceListResponse {
+  traces: TraceListItem[];
+}
+
 interface RawStateResponse {
   topology?: RawTopology;
   instances?: Record<
@@ -2555,6 +2585,49 @@ class AppStore {
       throw error;
     }
   }
+
+  /**
+   * List all available traces
+   */
+  async listTraces(): Promise<TraceListResponse> {
+    const response = await fetch("/v1/traces");
+    if (!response.ok) {
+      throw new Error(`Failed to list traces: ${response.status}`);
+    }
+    return (await response.json()) as TraceListResponse;
+  }
+
+  /**
+   * Check if a trace exists for a given task ID
+   */
+  async checkTraceExists(taskId: string): Promise<boolean> {
+    try {
+      const response = await fetch(`/v1/traces/${encodeURIComponent(taskId)}`);
+      return response.ok;
+    } catch {
+      return false;
+    }
+  }
+
+  /**
+   * Get computed statistics for a task's trace
+   */
+  async fetchTraceStats(taskId: string): Promise<TraceStatsResponse> {
+    const response = await fetch(
+      `/v1/traces/${encodeURIComponent(taskId)}/stats`,
+    );
+    if (!response.ok) {
+      throw new Error(`Failed to fetch trace stats: ${response.status}`);
+    }
+    return (await response.json()) as TraceStatsResponse;
+  }
+
+  /**
+   * Get the URL for the raw trace file (for Perfetto)
+   */
+  getTraceRawUrl(taskId: string): string {
+    return `/v1/traces/${encodeURIComponent(taskId)}/raw`;
+  }
 }
 
 export const appStore = new AppStore();
@@ -2666,3 +2739,12 @@ export const startDownload = (nodeId: string, shardMetadata: object) =>
   appStore.startDownload(nodeId, shardMetadata);
 export const deleteDownload = (nodeId: string, modelId: string) =>
   appStore.deleteDownload(nodeId, modelId);
+
+// Trace actions
+export const listTraces = () => appStore.listTraces();
+export const checkTraceExists = (taskId: string) =>
+  appStore.checkTraceExists(taskId);
+export const fetchTraceStats = (taskId: string) =>
+  appStore.fetchTraceStats(taskId);
+export const getTraceRawUrl = (taskId: string) =>
+  appStore.getTraceRawUrl(taskId);

dashboard/src/lib/stores/favorites.svelte.ts

@@ -0,0 +1,97 @@
+/**
+ * FavoritesStore - Manages favorite models with localStorage persistence
+ */
+
+import { browser } from "$app/environment";
+
+const FAVORITES_KEY = "exo-favorite-models";
+
+class FavoritesStore {
+  favorites = $state<Set<string>>(new Set());
+
+  constructor() {
+    if (browser) {
+      this.loadFromStorage();
+    }
+  }
+
+  private loadFromStorage() {
+    try {
+      const stored = localStorage.getItem(FAVORITES_KEY);
+      if (stored) {
+        const parsed = JSON.parse(stored) as string[];
+        this.favorites = new Set(parsed);
+      }
+    } catch (error) {
+      console.error("Failed to load favorites:", error);
+    }
+  }
+
+  private saveToStorage() {
+    try {
+      const array = Array.from(this.favorites);
+      localStorage.setItem(FAVORITES_KEY, JSON.stringify(array));
+    } catch (error) {
+      console.error("Failed to save favorites:", error);
+    }
+  }
+
+  add(baseModelId: string) {
+    const next = new Set(this.favorites);
+    next.add(baseModelId);
+    this.favorites = next;
+    this.saveToStorage();
+  }
+
+  remove(baseModelId: string) {
+    const next = new Set(this.favorites);
+    next.delete(baseModelId);
+    this.favorites = next;
+    this.saveToStorage();
+  }
+
+  toggle(baseModelId: string) {
+    if (this.favorites.has(baseModelId)) {
+      this.remove(baseModelId);
+    } else {
+      this.add(baseModelId);
+    }
+  }
+
+  isFavorite(baseModelId: string): boolean {
+    return this.favorites.has(baseModelId);
+  }
+
+  getAll(): string[] {
+    return Array.from(this.favorites);
+  }
+
+  getSet(): Set<string> {
+    return new Set(this.favorites);
+  }
+
+  hasAny(): boolean {
+    return this.favorites.size > 0;
+  }
+
+  clearAll() {
+    this.favorites = new Set();
+    this.saveToStorage();
+  }
+}
+
+export const favoritesStore = new FavoritesStore();
+
+export const favorites = () => favoritesStore.favorites;
+export const hasFavorites = () => favoritesStore.hasAny();
+export const isFavorite = (baseModelId: string) =>
+  favoritesStore.isFavorite(baseModelId);
+export const toggleFavorite = (baseModelId: string) =>
+  favoritesStore.toggle(baseModelId);
+export const addFavorite = (baseModelId: string) =>
+  favoritesStore.add(baseModelId);
+export const removeFavorite = (baseModelId: string) =>
+  favoritesStore.remove(baseModelId);
+export const getFavorites = () => favoritesStore.getAll();
+export const getFavoritesSet = () => favoritesStore.getSet();
+export const clearFavorites = () => favoritesStore.clearAll();

dashboard/src/routes/+page.svelte

@@ -5,7 +5,13 @@
     ChatMessages,
     ChatSidebar,
     ModelCard,
+    ModelPickerModal,
   } from "$lib/components";
+  import {
+    favorites,
+    toggleFavorite,
+    getFavoritesSet,
+  } from "$lib/stores/favorites.svelte";
   import {
     hasStartedChat,
     isTopologyMinimized,
@@ -100,6 +106,11 @@
       storage_size_megabytes?: number;
       tasks?: string[];
       hugging_face_id?: string;
+      is_custom?: boolean;
+      family?: string;
+      quantization?: string;
+      base_model?: string;
+      capabilities?: string[];
     }>
   >([]);
 
@@ -211,9 +222,11 @@
   let launchingModelId = $state<string | null>(null);
   let instanceDownloadExpandedNodes = $state<Set<string>>(new Set());
 
-  // Custom dropdown state
-  let isModelDropdownOpen = $state(false);
-  let modelDropdownSearch = $state("");
+  // Model picker modal state
+  let isModelPickerOpen = $state(false);
+
+  // Favorites state (reactive)
+  const favoritesSet = $derived(getFavoritesSet());
 
   // Slider dragging state
   let isDraggingSlider = $state(false);
@@ -530,6 +543,47 @@
     }
   }
 
+  async function addModelFromPicker(modelId: string) {
+    const response = await fetch("/models/add", {
+      method: "POST",
+      headers: { "Content-Type": "application/json" },
+      body: JSON.stringify({ model_id: modelId }),
+    });
+
+    if (!response.ok) {
+      let message = `Failed to add model (${response.status}: ${response.statusText})`;
+      try {
+        const err = await response.json();
+        if (err.detail) message = err.detail;
+      } catch {
+        // use default message
+      }
+      throw new Error(message);
+    }
+
+    await fetchModels();
+  }
+
+  async function deleteCustomModel(modelId: string) {
+    try {
+      const response = await fetch(
+        `/models/custom/${encodeURIComponent(modelId)}`,
+        { method: "DELETE" },
+      );
+      if (response.ok) {
+        await fetchModels();
+      }
+    } catch {
+      console.error("Failed to delete custom model");
+    }
+  }
+
+  function handleModelPickerSelect(modelId: string) {
+    selectPreviewModel(modelId);
+    saveLaunchDefaults();
+    isModelPickerOpen = false;
+  }
+
   async function launchInstance(
     modelId: string,
     specificPreview?: PlacementPreview | null,
@@ -2360,14 +2414,12 @@
               >
             </div>
 
-            <!-- Model Dropdown (Custom) -->
-            <div class="flex-shrink-0 mb-3 relative">
+            <!-- Model Picker Button -->
+            <div class="flex-shrink-0 mb-3">
               <button
                 type="button"
-                onclick={() => (isModelDropdownOpen = !isModelDropdownOpen)}
-                class="w-full bg-exo-medium-gray/50 border border-exo-yellow/30 rounded pl-3 pr-8 py-2.5 text-sm font-mono text-left tracking-wide cursor-pointer transition-all duration-200 hover:border-exo-yellow/50 focus:outline-none focus:border-exo-yellow/70 {isModelDropdownOpen
-                  ? 'border-exo-yellow/70'
-                  : ''}"
+                onclick={() => (isModelPickerOpen = true)}
+                class="w-full bg-exo-medium-gray/50 border border-exo-yellow/30 rounded pl-3 pr-8 py-2.5 text-sm font-mono text-left tracking-wide cursor-pointer transition-all duration-200 hover:border-exo-yellow/50 focus:outline-none focus:border-exo-yellow/70 relative"
               >
                 {#if selectedModelId}
                   {@const foundModel = models.find(
@@ -2375,54 +2427,12 @@
                   )}
                   {#if foundModel}
                     {@const sizeGB = getModelSizeGB(foundModel)}
-                    {@const isImageModel = modelSupportsImageGeneration(
-                      foundModel.id,
-                    )}
-                    {@const isImageEditModel = modelSupportsImageEditing(
-                      foundModel.id,
-                    )}
                     <span
                       class="flex items-center justify-between gap-2 w-full pr-4"
                     >
                       <span
                         class="flex items-center gap-2 text-exo-light-gray truncate"
                       >
-                        {#if isImageModel}
-                          <svg
-                            class="w-4 h-4 flex-shrink-0 text-exo-yellow"
-                            fill="none"
-                            viewBox="0 0 24 24"
-                            stroke="currentColor"
-                            stroke-width="2"
-                          >
-                            <rect
-                              x="3"
-                              y="3"
-                              width="18"
-                              height="18"
-                              rx="2"
-                              ry="2"
-                            />
-                            <circle cx="8.5" cy="8.5" r="1.5" />
-                            <polyline points="21 15 16 10 5 21" />
-                          </svg>
-                        {/if}
-                        {#if isImageEditModel}
-                          <svg
-                            class="w-4 h-4 flex-shrink-0 text-exo-yellow"
-                            fill="none"
-                            viewBox="0 0 24 24"
-                            stroke="currentColor"
-                            stroke-width="2"
-                          >
-                            <path
-                              d="M11 4H4a2 2 0 0 0-2 2v14a2 2 0 0 0 2 2h14a2 2 0 0 0 2-2v-7"
-                            />
-                            <path
-                              d="M18.5 2.5a2.121 2.121 0 0 1 3 3L12 15l-4 1 1-4 9.5-9.5z"
-                            />
-                          </svg>
-                        {/if}
                         <span class="truncate"
                           >{foundModel.name || foundModel.id}</span
                         >
@@ -2439,142 +2449,24 @@
                 {:else}
                   <span class="text-white/50">— SELECT MODEL —</span>
                 {/if}
-              </button>
-              <div
-                class="absolute right-3 top-1/2 -translate-y-1/2 pointer-events-none transition-transform duration-200 {isModelDropdownOpen
-                  ? 'rotate-180'
-                  : ''}"
-              >
-                <svg
-                  class="w-4 h-4 text-exo-yellow/60"
-                  fill="none"
-                  viewBox="0 0 24 24"
-                  stroke="currentColor"
-                >
-                  <path
-                    stroke-linecap="round"
-                    stroke-linejoin="round"
-                    stroke-width="2"
-                    d="M19 9l-7 7-7-7"
-                  />
-                </svg>
-              </div>
-
-              {#if isModelDropdownOpen}
-                <!-- Backdrop to close dropdown -->
-                <button
-                  type="button"
-                  class="fixed inset-0 z-40 cursor-default"
-                  onclick={() => (isModelDropdownOpen = false)}
-                  aria-label="Close dropdown"
-                ></button>
-
-                <!-- Dropdown Panel -->
                 <div
-                  class="absolute top-full left-0 right-0 mt-1 bg-exo-dark-gray border border-exo-yellow/30 rounded shadow-lg shadow-black/50 z-50 max-h-64 overflow-y-auto"
+                  class="absolute right-3 top-1/2 -translate-y-1/2 pointer-events-none"
                 >
-                  <!-- Search within dropdown -->
-                  <div
-                    class="sticky top-0 bg-exo-dark-gray border-b border-exo-medium-gray/30 p-2"
+                  <svg
+                    class="w-4 h-4 text-exo-yellow/60"
+                    fill="none"
+                    viewBox="0 0 24 24"
+                    stroke="currentColor"
                   >
-                    <input
-                      type="text"
-                      placeholder="Search models..."
-                      bind:value={modelDropdownSearch}
-                      class="w-full bg-exo-dark-gray/60 border border-exo-medium-gray/30 rounded px-2 py-1.5 text-xs font-mono text-white/80 placeholder:text-white/40 focus:outline-none focus:border-exo-yellow/50"
+                    <path
+                      stroke-linecap="round"
+                      stroke-linejoin="round"
+                      stroke-width="2"
+                      d="M19 9l-7 7-7-7"
                     />
-                  </div>
-
-                  <!-- Options -->
-                  <div class="py-1">
-                    {#each sortedModels().filter((m) => !modelDropdownSearch || (m.name || m.id)
-                          .toLowerCase()
-                          .includes(modelDropdownSearch.toLowerCase())) as model}
-                      {@const sizeGB = getModelSizeGB(model)}
-                      {@const modelCanFit = hasEnoughMemory(model)}
-                      {@const isImageModel = modelSupportsImageGeneration(
-                        model.id,
-                      )}
-                      {@const isImageEditModel = modelSupportsImageEditing(
-                        model.id,
-                      )}
-                      <button
-                        type="button"
-                        onclick={() => {
-                          if (modelCanFit) {
-                            selectPreviewModel(model.id);
-                            saveLaunchDefaults();
-                            isModelDropdownOpen = false;
-                            modelDropdownSearch = "";
-                          }
-                        }}
-                        disabled={!modelCanFit}
-                        class="w-full px-3 py-2 text-left text-sm font-mono tracking-wide transition-colors duration-100 flex items-center justify-between gap-2 {selectedModelId ===
-                        model.id
-                          ? 'bg-transparent text-exo-yellow cursor-pointer'
-                          : modelCanFit
-                            ? 'text-white/80 hover:text-exo-yellow cursor-pointer'
-                            : 'text-white/30 cursor-default'}"
-                      >
-                        <span class="flex items-center gap-2 truncate flex-1">
-                          {#if isImageModel}
-                            <svg
-                              class="w-4 h-4 flex-shrink-0 text-exo-yellow"
-                              fill="none"
-                              viewBox="0 0 24 24"
-                              stroke="currentColor"
-                              stroke-width="2"
-                              aria-label="Image generation model"
-                            >
-                              <rect
-                                x="3"
-                                y="3"
-                                width="18"
-                                height="18"
-                                rx="2"
-                                ry="2"
-                              />
-                              <circle cx="8.5" cy="8.5" r="1.5" />
-                              <polyline points="21 15 16 10 5 21" />
-                            </svg>
-                          {/if}
-                          {#if isImageEditModel}
-                            <svg
-                              class="w-4 h-4 flex-shrink-0 text-exo-yellow"
-                              fill="none"
-                              viewBox="0 0 24 24"
-                              stroke="currentColor"
-                              stroke-width="2"
-                              aria-label="Image editing model"
-                            >
-                              <path
-                                d="M11 4H4a2 2 0 0 0-2 2v14a2 2 0 0 0 2 2h14a2 2 0 0 0 2-2v-7"
-                              />
-                              <path
-                                d="M18.5 2.5a2.121 2.121 0 0 1 3 3L12 15l-4 1 1-4 9.5-9.5z"
-                              />
-                            </svg>
-                          {/if}
-                          <span class="truncate">{model.name || model.id}</span>
-                        </span>
-                        <span
-                          class="flex-shrink-0 text-xs {modelCanFit
-                            ? 'text-white/50'
-                            : 'text-red-400/60'}"
-                        >
-                          {sizeGB >= 1
-                            ? sizeGB.toFixed(0)
-                            : sizeGB.toFixed(1)}GB
-                        </span>
-                      </button>
-                    {:else}
-                      <div class="px-3 py-2 text-xs text-white/50 font-mono">
-                        No models found
-                      </div>
-                    {/each}
-                  </div>
+                  </svg>
                 </div>
-              {/if}
+              </button>
             </div>
 
             <!-- Configuration Options -->
@@ -3354,3 +3246,22 @@
     {/if}
   </main>
 </div>
+
+<ModelPickerModal
+  isOpen={isModelPickerOpen}
+  {models}
+  {selectedModelId}
+  favorites={favoritesSet}
+  existingModelIds={new Set(models.map((m) => m.id))}
+  canModelFit={(modelId) => {
+    const model = models.find((m) => m.id === modelId);
+    return model ? hasEnoughMemory(model) : false;
+  }}
+  onSelect={handleModelPickerSelect}
+  onClose={() => (isModelPickerOpen = false)}
+  onToggleFavorite={toggleFavorite}
+  onAddModel={addModelFromPicker}
+  onDeleteModel={deleteCustomModel}
+  totalMemoryGB={clusterMemory().total / (1024 * 1024 * 1024)}
+  usedMemoryGB={clusterMemory().used / (1024 * 1024 * 1024)}
+/>

dashboard/src/routes/traces/+page.svelte

@@ -0,0 +1,190 @@
+<script lang="ts">
+  import { onMount } from "svelte";
+  import {
+    listTraces,
+    getTraceRawUrl,
+    type TraceListItem,
+  } from "$lib/stores/app.svelte";
+  import HeaderNav from "$lib/components/HeaderNav.svelte";
+
+  let traces = $state<TraceListItem[]>([]);
+  let loading = $state(true);
+  let error = $state<string | null>(null);
+
+  function formatBytes(bytes: number): string {
+    if (!bytes || bytes <= 0) return "0B";
+    const units = ["B", "KB", "MB", "GB"];
+    const i = Math.min(
+      Math.floor(Math.log(bytes) / Math.log(1024)),
+      units.length - 1,
+    );
+    const val = bytes / Math.pow(1024, i);
+    return `${val.toFixed(val >= 10 ? 0 : 1)}${units[i]}`;
+  }
+
+  function formatDate(isoString: string): string {
+    const date = new Date(isoString);
+    return date.toLocaleString();
+  }
+
+  async function downloadTrace(taskId: string) {
+    const response = await fetch(getTraceRawUrl(taskId));
+    const blob = await response.blob();
+    const url = URL.createObjectURL(blob);
+    const a = document.createElement("a");
+    a.href = url;
+    a.download = `trace_${taskId}.json`;
+    a.click();
+    URL.revokeObjectURL(url);
+  }
+
+  async function openInPerfetto(taskId: string) {
+    // Fetch trace data from our local API
+    const response = await fetch(getTraceRawUrl(taskId));
+    const traceData = await response.arrayBuffer();
+
+    // Open Perfetto UI
+    const perfettoWindow = window.open("https://ui.perfetto.dev");
+    if (!perfettoWindow) {
+      alert("Failed to open Perfetto. Please allow popups.");
+      return;
+    }
+
+    // Wait for Perfetto to be ready, then send trace via postMessage
+    const onMessage = (e: MessageEvent) => {
+      if (e.data === "PONG") {
+        window.removeEventListener("message", onMessage);
+        perfettoWindow.postMessage(
+          {
+            perfetto: {
+              buffer: traceData,
+              title: `Trace ${taskId}`,
+            },
+          },
+          "https://ui.perfetto.dev",
+        );
+      }
+    };
+    window.addEventListener("message", onMessage);
+
+    // Ping Perfetto until it responds
+    const pingInterval = setInterval(() => {
+      perfettoWindow.postMessage("PING", "https://ui.perfetto.dev");
+    }, 50);
+
+    // Clean up after 10 seconds
+    setTimeout(() => {
+      clearInterval(pingInterval);
+      window.removeEventListener("message", onMessage);
+    }, 10000);
+  }
+
+  async function refresh() {
+    loading = true;
+    error = null;
+    try {
+      const response = await listTraces();
+      traces = response.traces;
+    } catch (e) {
+      error = e instanceof Error ? e.message : "Failed to load traces";
+    } finally {
+      loading = false;
+    }
+  }
+
+  onMount(() => {
+    refresh();
+  });
+</script>
+
+<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">
+    <div class="flex items-center justify-between gap-4 flex-wrap">
+      <div>
+        <h1
+          class="text-2xl font-mono tracking-[0.2em] uppercase text-exo-yellow"
+        >
+          Traces
+        </h1>
+      </div>
+      <div class="flex items-center gap-3">
+        <button
+          type="button"
+          class="text-xs font-mono text-exo-light-gray hover:text-exo-yellow transition-colors uppercase border border-exo-medium-gray/40 px-2 py-1 rounded"
+          onclick={refresh}
+          disabled={loading}
+        >
+          Refresh
+        </button>
+      </div>
+    </div>
+
+    {#if loading}
+      <div
+        class="rounded border border-exo-medium-gray/30 bg-exo-black/30 p-6 text-center text-exo-light-gray"
+      >
+        <div class="text-sm">Loading traces...</div>
+      </div>
+    {:else if error}
+      <div
+        class="rounded border border-red-500/30 bg-red-500/10 p-6 text-center text-red-400"
+      >
+        <div class="text-sm">{error}</div>
+      </div>
+    {:else if traces.length === 0}
+      <div
+        class="rounded border border-exo-medium-gray/30 bg-exo-black/30 p-6 text-center text-exo-light-gray space-y-2"
+      >
+        <div class="text-sm">No traces found.</div>
+        <div class="text-xs text-exo-light-gray/70">
+          Run exo with EXO_TRACING_ENABLED=1 to collect traces.
+        </div>
+      </div>
+    {:else}
+      <div class="space-y-3">
+        {#each traces as trace}
+          <div
+            class="rounded border border-exo-medium-gray/30 bg-exo-black/30 p-4 flex items-center justify-between gap-4"
+          >
+            <div class="min-w-0 flex-1">
+              <a
+                href="#/traces/{trace.taskId}"
+                class="text-sm font-mono text-white hover:text-exo-yellow transition-colors truncate block"
+              >
+                {trace.taskId}
+              </a>
+              <div class="text-xs text-exo-light-gray font-mono mt-1">
+                {formatDate(trace.createdAt)} &bull; {formatBytes(
+                  trace.fileSize,
+                )}
+              </div>
+            </div>
+            <div class="flex items-center gap-2 shrink-0">
+              <a
+                href="#/traces/{trace.taskId}"
+                class="text-xs font-mono text-exo-light-gray hover:text-exo-yellow transition-colors uppercase border border-exo-medium-gray/40 px-2 py-1 rounded"
+              >
+                View Stats
+              </a>
+              <button
+                type="button"
+                class="text-xs font-mono text-exo-light-gray hover:text-exo-yellow transition-colors uppercase border border-exo-medium-gray/40 px-2 py-1 rounded"
+                onclick={() => downloadTrace(trace.taskId)}
+              >
+                Download
+              </button>
+              <button
+                type="button"
+                class="text-xs font-mono text-exo-dark-gray bg-exo-yellow hover:bg-exo-yellow/90 transition-colors uppercase px-2 py-1 rounded font-semibold"
+                onclick={() => openInPerfetto(trace.taskId)}
+              >
+                View Trace
+              </button>
+            </div>
+          </div>
+        {/each}
+      </div>
+    {/if}
+  </div>
+</div>

dashboard/src/routes/traces/[taskId]/+page.svelte

@@ -0,0 +1,367 @@
+<script lang="ts">
+  import { page } from "$app/stores";
+  import { onMount } from "svelte";
+  import {
+    fetchTraceStats,
+    getTraceRawUrl,
+    type TraceStatsResponse,
+    type TraceCategoryStats,
+  } from "$lib/stores/app.svelte";
+  import HeaderNav from "$lib/components/HeaderNav.svelte";
+
+  const taskId = $derived($page.params.taskId);
+
+  let stats = $state<TraceStatsResponse | null>(null);
+  let loading = $state(true);
+  let error = $state<string | null>(null);
+
+  function formatDuration(us: number): string {
+    if (us < 1000) return `${us.toFixed(0)}us`;
+    if (us < 1_000_000) return `${(us / 1000).toFixed(2)}ms`;
+    return `${(us / 1_000_000).toFixed(2)}s`;
+  }
+
+  function formatPercentage(part: number, total: number): string {
+    if (total === 0) return "0.0%";
+    return `${((part / total) * 100).toFixed(1)}%`;
+  }
+
+  // Parse hierarchical categories like "sync/compute" into phases
+  type PhaseData = {
+    name: string;
+    subcategories: { name: string; stats: TraceCategoryStats }[];
+    totalUs: number; // From outer span (e.g., "sync" category)
+    stepCount: number; // Count of outer span events
+  };
+
+  function parsePhases(
+    byCategory: Record<string, TraceCategoryStats>,
+  ): PhaseData[] {
+    const phases = new Map<
+      string,
+      {
+        subcats: Map<string, TraceCategoryStats>;
+        outerStats: TraceCategoryStats | null;
+      }
+    >();
+
+    for (const [category, catStats] of Object.entries(byCategory)) {
+      if (category.includes("/")) {
+        const [phase, subcat] = category.split("/", 2);
+        if (!phases.has(phase)) {
+          phases.set(phase, { subcats: new Map(), outerStats: null });
+        }
+        phases.get(phase)!.subcats.set(subcat, catStats);
+      } else {
+        // Outer span - this IS the phase total
+        if (!phases.has(category)) {
+          phases.set(category, { subcats: new Map(), outerStats: null });
+        }
+        phases.get(category)!.outerStats = catStats;
+      }
+    }
+
+    return Array.from(phases.entries())
+      .filter(([_, data]) => data.outerStats !== null) // Only phases with outer spans
+      .map(([name, data]) => ({
+        name,
+        subcategories: Array.from(data.subcats.entries())
+          .map(([subName, subStats]) => ({ name: subName, stats: subStats }))
+          .sort((a, b) => b.stats.totalUs - a.stats.totalUs),
+        totalUs: data.outerStats!.totalUs, // Outer span total
+        stepCount: data.outerStats!.count, // Number of steps
+      }))
+      .sort((a, b) => b.totalUs - a.totalUs);
+  }
+
+  async function downloadTrace() {
+    if (!taskId) return;
+    const response = await fetch(getTraceRawUrl(taskId));
+    const blob = await response.blob();
+    const url = URL.createObjectURL(blob);
+    const a = document.createElement("a");
+    a.href = url;
+    a.download = `trace_${taskId}.json`;
+    a.click();
+    URL.revokeObjectURL(url);
+  }
+
+  async function openInPerfetto() {
+    if (!taskId) return;
+
+    // Fetch trace data from our local API
+    const response = await fetch(getTraceRawUrl(taskId));
+    const traceData = await response.arrayBuffer();
+
+    // Open Perfetto UI
+    const perfettoWindow = window.open("https://ui.perfetto.dev");
+    if (!perfettoWindow) {
+      alert("Failed to open Perfetto. Please allow popups.");
+      return;
+    }
+
+    // Wait for Perfetto to be ready, then send trace via postMessage
+    const onMessage = (e: MessageEvent) => {
+      if (e.data === "PONG") {
+        window.removeEventListener("message", onMessage);
+        perfettoWindow.postMessage(
+          {
+            perfetto: {
+              buffer: traceData,
+              title: `Trace ${taskId}`,
+            },
+          },
+          "https://ui.perfetto.dev",
+        );
+      }
+    };
+    window.addEventListener("message", onMessage);
+
+    // Ping Perfetto until it responds
+    const pingInterval = setInterval(() => {
+      perfettoWindow.postMessage("PING", "https://ui.perfetto.dev");
+    }, 50);
+
+    // Clean up after 10 seconds
+    setTimeout(() => {
+      clearInterval(pingInterval);
+      window.removeEventListener("message", onMessage);
+    }, 10000);
+  }
+
+  onMount(async () => {
+    if (!taskId) {
+      error = "No task ID provided";
+      loading = false;
+      return;
+    }
+
+    try {
+      stats = await fetchTraceStats(taskId);
+    } catch (e) {
+      error = e instanceof Error ? e.message : "Failed to load trace";
+    } finally {
+      loading = false;
+    }
+  });
+
+  const phases = $derived(stats ? parsePhases(stats.byCategory) : []);
+  const sortedRanks = $derived(
+    stats
+      ? Object.keys(stats.byRank)
+          .map(Number)
+          .sort((a, b) => a - b)
+      : [],
+  );
+  const nodeCount = $derived(sortedRanks.length || 1);
+</script>
+
+<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">
+    <div class="flex items-center justify-between gap-4 flex-wrap">
+      <div>
+        <h1
+          class="text-2xl font-mono tracking-[0.2em] uppercase text-exo-yellow"
+        >
+          Trace
+        </h1>
+        <p class="text-sm text-exo-light-gray font-mono truncate max-w-lg">
+          {taskId}
+        </p>
+      </div>
+      <div class="flex items-center gap-3">
+        <a
+          href="#/traces"
+          class="text-xs font-mono text-exo-light-gray hover:text-exo-yellow transition-colors uppercase border border-exo-medium-gray/40 px-3 py-1.5 rounded"
+        >
+          All Traces
+        </a>
+        <button
+          type="button"
+          class="text-xs font-mono text-exo-light-gray hover:text-exo-yellow transition-colors uppercase border border-exo-medium-gray/40 px-3 py-1.5 rounded"
+          onclick={downloadTrace}
+          disabled={loading || !!error}
+        >
+          Download
+        </button>
+        <button
+          type="button"
+          class="text-xs font-mono text-exo-dark-gray bg-exo-yellow hover:bg-exo-yellow/90 transition-colors uppercase px-3 py-1.5 rounded font-semibold"
+          onclick={openInPerfetto}
+          disabled={loading || !!error}
+        >
+          View Trace
+        </button>
+      </div>
+    </div>
+
+    {#if loading}
+      <div
+        class="rounded border border-exo-medium-gray/30 bg-exo-black/30 p-6 text-center text-exo-light-gray"
+      >
+        <div class="text-sm">Loading trace data...</div>
+      </div>
+    {:else if error}
+      <div
+        class="rounded border border-red-500/30 bg-red-500/10 p-6 text-center text-red-400"
+      >
+        <div class="text-sm">{error}</div>
+      </div>
+    {:else if stats}
+      <!-- Wall Time Summary -->
+      <div
+        class="rounded border border-exo-medium-gray/30 bg-exo-black/30 p-4 space-y-2"
+      >
+        <h2
+          class="text-sm font-mono uppercase tracking-wider text-exo-light-gray"
+        >
+          Summary
+        </h2>
+        <div class="text-3xl font-mono text-exo-yellow">
+          {formatDuration(stats.totalWallTimeUs)}
+        </div>
+        <div class="text-xs text-exo-light-gray">Total wall time</div>
+      </div>
+
+      <!-- By Phase -->
+      {#if phases.length > 0}
+        <div
+          class="rounded border border-exo-medium-gray/30 bg-exo-black/30 p-4 space-y-4"
+        >
+          <h2
+            class="text-sm font-mono uppercase tracking-wider text-exo-light-gray"
+          >
+            By Phase <span class="text-exo-light-gray/50">(avg per node)</span>
+          </h2>
+          <div class="space-y-4">
+            {#each phases as phase}
+              {@const normalizedTotal = phase.totalUs / nodeCount}
+              {@const normalizedStepCount = phase.stepCount / nodeCount}
+              <div class="space-y-2">
+                <div class="flex items-center justify-between">
+                  <span class="text-sm font-mono text-white">{phase.name}</span>
+                  <span class="text-sm font-mono">
+                    <span class="text-exo-yellow"
+                      >{formatDuration(normalizedTotal)}</span
+                    >
+                    <span class="text-exo-light-gray ml-2">
+                      ({normalizedStepCount} steps, {formatDuration(
+                        normalizedTotal / normalizedStepCount,
+                      )}/step)
+                    </span>
+                  </span>
+                </div>
+                {#if phase.subcategories.length > 0}
+                  <div class="pl-4 space-y-1.5">
+                    {#each phase.subcategories as subcat}
+                      {@const normalizedSubcat =
+                        subcat.stats.totalUs / nodeCount}
+                      {@const pct = formatPercentage(
+                        normalizedSubcat,
+                        normalizedTotal,
+                      )}
+                      {@const perStep = normalizedSubcat / normalizedStepCount}
+                      <div
+                        class="flex items-center justify-between text-xs font-mono"
+                      >
+                        <span class="text-exo-light-gray">{subcat.name}</span>
+                        <span class="text-white">
+                          {formatDuration(normalizedSubcat)}
+                          <span class="text-exo-light-gray ml-2">({pct})</span>
+                          <span class="text-exo-light-gray/60 ml-2"
+                            >{formatDuration(perStep)}/step</span
+                          >
+                        </span>
+                      </div>
+                      <!-- Progress bar -->
+                      <div
+                        class="relative h-1.5 bg-exo-black/60 rounded-sm overflow-hidden"
+                      >
+                        <div
+                          class="absolute inset-y-0 left-0 bg-gradient-to-r from-exo-yellow to-exo-yellow/70 transition-all duration-300"
+                          style="width: {pct}"
+                        ></div>
+                      </div>
+                    {/each}
+                  </div>
+                {/if}
+              </div>
+            {/each}
+          </div>
+        </div>
+      {/if}
+
+      <!-- By Rank -->
+      {#if sortedRanks.length > 0}
+        <div
+          class="rounded border border-exo-medium-gray/30 bg-exo-black/30 p-4 space-y-4"
+        >
+          <h2
+            class="text-sm font-mono uppercase tracking-wider text-exo-light-gray"
+          >
+            By Rank
+          </h2>
+          <div class="grid grid-cols-1 md:grid-cols-2 lg:grid-cols-3 gap-4">
+            {#each sortedRanks as rank}
+              {@const rankStats = stats.byRank[rank]}
+              {@const rankPhases = parsePhases(rankStats.byCategory)}
+              <div
+                class="rounded border border-exo-medium-gray/20 bg-exo-dark-gray/60 p-3 space-y-3"
+              >
+                <div class="text-sm font-mono text-exo-yellow">
+                  Rank {rank}
+                </div>
+                <div class="space-y-2">
+                  {#each rankPhases as phase}
+                    <div class="space-y-1">
+                      <div class="flex items-center justify-between text-xs">
+                        <span class="font-mono text-exo-light-gray"
+                          >{phase.name}</span
+                        >
+                        <span class="font-mono text-white">
+                          {formatDuration(phase.totalUs)}
+                          <span class="text-exo-light-gray/50 ml-1">
+                            ({phase.stepCount}x)
+                          </span>
+                        </span>
+                      </div>
+                      {#if phase.subcategories.length > 0}
+                        <div class="pl-2 space-y-0.5">
+                          {#each phase.subcategories as subcat}
+                            {@const pct = formatPercentage(
+                              subcat.stats.totalUs,
+                              phase.totalUs,
+                            )}
+                            {@const perStep =
+                              subcat.stats.totalUs / phase.stepCount}
+                            <div
+                              class="flex items-center justify-between text-[10px] font-mono"
+                            >
+                              <span class="text-exo-light-gray/70"
+                                >{subcat.name}</span
+                              >
+                              <span class="text-exo-light-gray">
+                                {formatDuration(subcat.stats.totalUs)}
+                                <span class="text-exo-light-gray/50"
+                                  >({pct})</span
+                                >
+                                <span class="text-exo-light-gray/30 ml-1"
+                                  >{formatDuration(perStep)}/step</span
+                                >
+                              </span>
+                            </div>
+                          {/each}
+                        </div>
+                      {/if}
+                    </div>
+                  {/each}
+                </div>
+              </div>
+            {/each}
+          </div>
+        </div>
+      {/if}
+    {/if}
+  </div>
+</div>

packaging/pyinstaller/exo.spec

@@ -10,6 +10,7 @@ PROJECT_ROOT = Path.cwd()
 SOURCE_ROOT = PROJECT_ROOT / "src"
 ENTRYPOINT = SOURCE_ROOT / "exo" / "__main__.py"
 DASHBOARD_DIR = PROJECT_ROOT / "dashboard" / "build"
+RESOURCES_DIR = PROJECT_ROOT / "resources"
 EXO_SHARED_MODELS_DIR = SOURCE_ROOT / "exo" / "shared" / "models"
 
 if not ENTRYPOINT.is_file():
@@ -18,6 +19,9 @@ if not ENTRYPOINT.is_file():
 if not DASHBOARD_DIR.is_dir():
     raise SystemExit(f"Dashboard assets are missing: {DASHBOARD_DIR}")
 
+if not RESOURCES_DIR.is_dir():
+    raise SystemExit(f"Resource assets are missing: {RESOURCES_DIR}")
+
 if not EXO_SHARED_MODELS_DIR.is_dir():
     raise SystemExit(f"Shared model assets are missing: {EXO_SHARED_MODELS_DIR}")
 
@@ -58,6 +62,7 @@ HIDDEN_IMPORTS = sorted(
 
 DATAS: list[tuple[str, str]] = [
     (str(DASHBOARD_DIR), "dashboard"),
+    (str(RESOURCES_DIR), "resources"),
     (str(MLX_LIB_DIR), "mlx/lib"),
     (str(EXO_SHARED_MODELS_DIR), "exo/shared/models"),
 ]

python/parts.nix

@@ -69,7 +69,9 @@
           # Create wrapper scripts
           for script in exo exo-master exo-worker; do
             makeWrapper ${exoVenv}/bin/$script $out/bin/$script \
-              --set DASHBOARD_DIR ${self'.packages.dashboard}
+              --set EXO_DASHBOARD_DIR ${self'.packages.dashboard} \
+              --set EXO_RESOURCES_DIR ${inputs.self + "/resources"} \
+              ${lib.optionalString pkgs.stdenv.isDarwin "--prefix PATH : ${pkgs.macmon}/bin"}
           done
         '';
     in

resources/image_model_cards/exolabs--FLUX.1-Krea-dev-4bit.toml

@@ -0,0 +1,45 @@
+model_id = "exolabs/FLUX.1-Krea-dev-4bit"
+n_layers = 57
+hidden_size = 1
+supports_tensor = false
+tasks = ["TextToImage"]
+
+[storage_size]
+in_bytes = 15475325472
+
+[[components]]
+component_name = "text_encoder"
+component_path = "text_encoder/"
+n_layers = 12
+can_shard = false
+
+[components.storage_size]
+in_bytes = 0
+
+[[components]]
+component_name = "text_encoder_2"
+component_path = "text_encoder_2/"
+n_layers = 24
+can_shard = false
+safetensors_index_filename = "model.safetensors.index.json"
+
+[components.storage_size]
+in_bytes = 9524621312
+
+[[components]]
+component_name = "transformer"
+component_path = "transformer/"
+n_layers = 57
+can_shard = true
+safetensors_index_filename = "diffusion_pytorch_model.safetensors.index.json"
+
+[components.storage_size]
+in_bytes = 5950704160
+
+[[components]]
+component_name = "vae"
+component_path = "vae/"
+can_shard = false
+
+[components.storage_size]
+in_bytes = 0

resources/image_model_cards/exolabs--FLUX.1-Krea-dev-8bit.toml

@@ -0,0 +1,45 @@
+model_id = "exolabs/FLUX.1-Krea-dev-8bit"
+n_layers = 57
+hidden_size = 1
+supports_tensor = false
+tasks = ["TextToImage"]
+
+[storage_size]
+in_bytes = 21426029632
+
+[[components]]
+component_name = "text_encoder"
+component_path = "text_encoder/"
+n_layers = 12
+can_shard = false
+
+[components.storage_size]
+in_bytes = 0
+
+[[components]]
+component_name = "text_encoder_2"
+component_path = "text_encoder_2/"
+n_layers = 24
+can_shard = false
+safetensors_index_filename = "model.safetensors.index.json"
+
+[components.storage_size]
+in_bytes = 9524621312
+
+[[components]]
+component_name = "transformer"
+component_path = "transformer/"
+n_layers = 57
+can_shard = true
+safetensors_index_filename = "diffusion_pytorch_model.safetensors.index.json"
+
+[components.storage_size]
+in_bytes = 11901408320
+
+[[components]]
+component_name = "vae"
+component_path = "vae/"
+can_shard = false
+
+[components.storage_size]
+in_bytes = 0

resources/image_model_cards/exolabs--FLUX.1-Krea-dev.toml

@@ -0,0 +1,45 @@
+model_id = "exolabs/FLUX.1-Krea-dev"
+n_layers = 57
+hidden_size = 1
+supports_tensor = false
+tasks = ["TextToImage"]
+
+[storage_size]
+in_bytes = 33327437952
+
+[[components]]
+component_name = "text_encoder"
+component_path = "text_encoder/"
+n_layers = 12
+can_shard = false
+
+[components.storage_size]
+in_bytes = 0
+
+[[components]]
+component_name = "text_encoder_2"
+component_path = "text_encoder_2/"
+n_layers = 24
+can_shard = false
+safetensors_index_filename = "model.safetensors.index.json"
+
+[components.storage_size]
+in_bytes = 9524621312
+
+[[components]]
+component_name = "transformer"
+component_path = "transformer/"
+n_layers = 57
+can_shard = true
+safetensors_index_filename = "diffusion_pytorch_model.safetensors.index.json"
+
+[components.storage_size]
+in_bytes = 23802816640
+
+[[components]]
+component_name = "vae"
+component_path = "vae/"
+can_shard = false
+
+[components.storage_size]
+in_bytes = 0

resources/image_model_cards/exolabs--FLUX.1-dev-4bit.toml

@@ -0,0 +1,45 @@
+model_id = "exolabs/FLUX.1-dev-4bit"
+n_layers = 57
+hidden_size = 1
+supports_tensor = false
+tasks = ["TextToImage"]
+
+[storage_size]
+in_bytes = 15475325472
+
+[[components]]
+component_name = "text_encoder"
+component_path = "text_encoder/"
+n_layers = 12
+can_shard = false
+
+[components.storage_size]
+in_bytes = 0
+
+[[components]]
+component_name = "text_encoder_2"
+component_path = "text_encoder_2/"
+n_layers = 24
+can_shard = false
+safetensors_index_filename = "model.safetensors.index.json"
+
+[components.storage_size]
+in_bytes = 9524621312
+
+[[components]]
+component_name = "transformer"
+component_path = "transformer/"
+n_layers = 57
+can_shard = true
+safetensors_index_filename = "diffusion_pytorch_model.safetensors.index.json"
+
+[components.storage_size]
+in_bytes = 5950704160
+
+[[components]]
+component_name = "vae"
+component_path = "vae/"
+can_shard = false
+
+[components.storage_size]
+in_bytes = 0

resources/image_model_cards/exolabs--FLUX.1-dev-8bit.toml

@@ -0,0 +1,45 @@
+model_id = "exolabs/FLUX.1-dev-8bit"
+n_layers = 57
+hidden_size = 1
+supports_tensor = false
+tasks = ["TextToImage"]
+
+[storage_size]
+in_bytes = 21426029632
+
+[[components]]
+component_name = "text_encoder"
+component_path = "text_encoder/"
+n_layers = 12
+can_shard = false
+
+[components.storage_size]
+in_bytes = 0
+
+[[components]]
+component_name = "text_encoder_2"
+component_path = "text_encoder_2/"
+n_layers = 24
+can_shard = false
+safetensors_index_filename = "model.safetensors.index.json"
+
+[components.storage_size]
+in_bytes = 9524621312
+
+[[components]]
+component_name = "transformer"
+component_path = "transformer/"
+n_layers = 57
+can_shard = true
+safetensors_index_filename = "diffusion_pytorch_model.safetensors.index.json"
+
+[components.storage_size]
+in_bytes = 11901408320
+
+[[components]]
+component_name = "vae"
+component_path = "vae/"
+can_shard = false
+
+[components.storage_size]
+in_bytes = 0

resources/image_model_cards/exolabs--FLUX.1-dev.toml

@@ -0,0 +1,45 @@
+model_id = "exolabs/FLUX.1-dev"
+n_layers = 57
+hidden_size = 1
+supports_tensor = false
+tasks = ["TextToImage"]
+
+[storage_size]
+in_bytes = 33327437952
+
+[[components]]
+component_name = "text_encoder"
+component_path = "text_encoder/"
+n_layers = 12
+can_shard = false
+
+[components.storage_size]
+in_bytes = 0
+
+[[components]]
+component_name = "text_encoder_2"
+component_path = "text_encoder_2/"
+n_layers = 24
+can_shard = false
+safetensors_index_filename = "model.safetensors.index.json"
+
+[components.storage_size]
+in_bytes = 9524621312
+
+[[components]]
+component_name = "transformer"
+component_path = "transformer/"
+n_layers = 57
+can_shard = true
+safetensors_index_filename = "diffusion_pytorch_model.safetensors.index.json"
+
+[components.storage_size]
+in_bytes = 23802816640
+
+[[components]]
+component_name = "vae"
+component_path = "vae/"
+can_shard = false
+
+[components.storage_size]
+in_bytes = 0

resources/image_model_cards/exolabs--FLUX.1-schnell-4bit.toml

@@ -0,0 +1,45 @@
+model_id = "exolabs/FLUX.1-schnell-4bit"
+n_layers = 57
+hidden_size = 1
+supports_tensor = false
+tasks = ["TextToImage"]
+
+[storage_size]
+in_bytes = 15470210592
+
+[[components]]
+component_name = "text_encoder"
+component_path = "text_encoder/"
+n_layers = 12
+can_shard = false
+
+[components.storage_size]
+in_bytes = 0
+
+[[components]]
+component_name = "text_encoder_2"
+component_path = "text_encoder_2/"
+n_layers = 24
+can_shard = false
+safetensors_index_filename = "model.safetensors.index.json"
+
+[components.storage_size]
+in_bytes = 9524621312
+
+[[components]]
+component_name = "transformer"
+component_path = "transformer/"
+n_layers = 57
+can_shard = true
+safetensors_index_filename = "diffusion_pytorch_model.safetensors.index.json"
+
+[components.storage_size]
+in_bytes = 5945589280
+
+[[components]]
+component_name = "vae"
+component_path = "vae/"
+can_shard = false
+
+[components.storage_size]
+in_bytes = 0

resources/image_model_cards/exolabs--FLUX.1-schnell-8bit.toml

@@ -0,0 +1,45 @@
+model_id = "exolabs/FLUX.1-schnell-8bit"
+n_layers = 57
+hidden_size = 1
+supports_tensor = false
+tasks = ["TextToImage"]
+
+[storage_size]
+in_bytes = 21415799872
+
+[[components]]
+component_name = "text_encoder"
+component_path = "text_encoder/"
+n_layers = 12
+can_shard = false
+
+[components.storage_size]
+in_bytes = 0
+
+[[components]]
+component_name = "text_encoder_2"
+component_path = "text_encoder_2/"
+n_layers = 24
+can_shard = false
+safetensors_index_filename = "model.safetensors.index.json"
+
+[components.storage_size]
+in_bytes = 9524621312
+
+[[components]]
+component_name = "transformer"
+component_path = "transformer/"
+n_layers = 57
+can_shard = true
+safetensors_index_filename = "diffusion_pytorch_model.safetensors.index.json"
+
+[components.storage_size]
+in_bytes = 11891178560
+
+[[components]]
+component_name = "vae"
+component_path = "vae/"
+can_shard = false
+
+[components.storage_size]
+in_bytes = 0

resources/image_model_cards/exolabs--FLUX.1-schnell.toml

@@ -0,0 +1,45 @@
+model_id = "exolabs/FLUX.1-schnell"
+n_layers = 57
+hidden_size = 1
+supports_tensor = false
+tasks = ["TextToImage"]
+
+[storage_size]
+in_bytes = 33306978432
+
+[[components]]
+component_name = "text_encoder"
+component_path = "text_encoder/"
+n_layers = 12
+can_shard = false
+
+[components.storage_size]
+in_bytes = 0
+
+[[components]]
+component_name = "text_encoder_2"
+component_path = "text_encoder_2/"
+n_layers = 24
+can_shard = false
+safetensors_index_filename = "model.safetensors.index.json"
+
+[components.storage_size]
+in_bytes = 9524621312
+
+[[components]]
+component_name = "transformer"
+component_path = "transformer/"
+n_layers = 57
+can_shard = true
+safetensors_index_filename = "diffusion_pytorch_model.safetensors.index.json"
+
+[components.storage_size]
+in_bytes = 23782357120
+
+[[components]]
+component_name = "vae"
+component_path = "vae/"
+can_shard = false
+
+[components.storage_size]
+in_bytes = 0

resources/image_model_cards/exolabs--Qwen-Image-4bit.toml

@@ -0,0 +1,36 @@
+model_id = "exolabs/Qwen-Image-4bit"
+n_layers = 60
+hidden_size = 1
+supports_tensor = false
+tasks = ["TextToImage"]
+uses_cfg = true
+
+[storage_size]
+in_bytes = 26799533856
+
+[[components]]
+component_name = "text_encoder"
+component_path = "text_encoder/"
+n_layers = 12
+can_shard = false
+
+[components.storage_size]
+in_bytes = 16584333312
+
+[[components]]
+component_name = "transformer"
+component_path = "transformer/"
+n_layers = 60
+can_shard = true
+safetensors_index_filename = "diffusion_pytorch_model.safetensors.index.json"
+
+[components.storage_size]
+in_bytes = 10215200544
+
+[[components]]
+component_name = "vae"
+component_path = "vae/"
+can_shard = false
+
+[components.storage_size]
+in_bytes = 0

resources/image_model_cards/exolabs--Qwen-Image-8bit.toml

@@ -0,0 +1,36 @@
+model_id = "exolabs/Qwen-Image-8bit"
+n_layers = 60
+hidden_size = 1
+supports_tensor = false
+tasks = ["TextToImage"]
+uses_cfg = true
+
+[storage_size]
+in_bytes = 37014734400
+
+[[components]]
+component_name = "text_encoder"
+component_path = "text_encoder/"
+n_layers = 12
+can_shard = false
+
+[components.storage_size]
+in_bytes = 16584333312
+
+[[components]]
+component_name = "transformer"
+component_path = "transformer/"
+n_layers = 60
+can_shard = true
+safetensors_index_filename = "diffusion_pytorch_model.safetensors.index.json"
+
+[components.storage_size]
+in_bytes = 20430401088
+
+[[components]]
+component_name = "vae"
+component_path = "vae/"
+can_shard = false
+
+[components.storage_size]
+in_bytes = 0

resources/image_model_cards/exolabs--Qwen-Image-Edit-2509-4bit.toml

@@ -0,0 +1,36 @@
+model_id = "exolabs/Qwen-Image-Edit-2509-4bit"
+n_layers = 60
+hidden_size = 1
+supports_tensor = false
+tasks = ["ImageToImage"]
+uses_cfg = true
+
+[storage_size]
+in_bytes = 26799533856
+
+[[components]]
+component_name = "text_encoder"
+component_path = "text_encoder/"
+n_layers = 12
+can_shard = false
+
+[components.storage_size]
+in_bytes = 16584333312
+
+[[components]]
+component_name = "transformer"
+component_path = "transformer/"
+n_layers = 60
+can_shard = true
+safetensors_index_filename = "diffusion_pytorch_model.safetensors.index.json"
+
+[components.storage_size]
+in_bytes = 10215200544
+
+[[components]]
+component_name = "vae"
+component_path = "vae/"
+can_shard = false
+
+[components.storage_size]
+in_bytes = 0

resources/image_model_cards/exolabs--Qwen-Image-Edit-2509-8bit.toml

@@ -0,0 +1,36 @@
+model_id = "exolabs/Qwen-Image-Edit-2509-8bit"
+n_layers = 60
+hidden_size = 1
+supports_tensor = false
+tasks = ["ImageToImage"]
+uses_cfg = true
+
+[storage_size]
+in_bytes = 37014734400
+
+[[components]]
+component_name = "text_encoder"
+component_path = "text_encoder/"
+n_layers = 12
+can_shard = false
+
+[components.storage_size]
+in_bytes = 16584333312
+
+[[components]]
+component_name = "transformer"
+component_path = "transformer/"
+n_layers = 60
+can_shard = true
+safetensors_index_filename = "diffusion_pytorch_model.safetensors.index.json"
+
+[components.storage_size]
+in_bytes = 20430401088
+
+[[components]]
+component_name = "vae"
+component_path = "vae/"
+can_shard = false
+
+[components.storage_size]
+in_bytes = 0

resources/image_model_cards/exolabs--Qwen-Image-Edit-2509.toml

@@ -0,0 +1,36 @@
+model_id = "exolabs/Qwen-Image-Edit-2509"
+n_layers = 60
+hidden_size = 1
+supports_tensor = false
+tasks = ["ImageToImage"]
+uses_cfg = true
+
+[storage_size]
+in_bytes = 57445135488
+
+[[components]]
+component_name = "text_encoder"
+component_path = "text_encoder/"
+n_layers = 12
+can_shard = false
+
+[components.storage_size]
+in_bytes = 16584333312
+
+[[components]]
+component_name = "transformer"
+component_path = "transformer/"
+n_layers = 60
+can_shard = true
+safetensors_index_filename = "diffusion_pytorch_model.safetensors.index.json"
+
+[components.storage_size]
+in_bytes = 40860802176
+
+[[components]]
+component_name = "vae"
+component_path = "vae/"
+can_shard = false
+
+[components.storage_size]
+in_bytes = 0

resources/image_model_cards/exolabs--Qwen-Image.toml

@@ -0,0 +1,36 @@
+model_id = "exolabs/Qwen-Image"
+n_layers = 60
+hidden_size = 1
+supports_tensor = false
+tasks = ["TextToImage"]
+uses_cfg = true
+
+[storage_size]
+in_bytes = 57445135488
+
+[[components]]
+component_name = "text_encoder"
+component_path = "text_encoder/"
+n_layers = 12
+can_shard = false
+
+[components.storage_size]
+in_bytes = 16584333312
+
+[[components]]
+component_name = "transformer"
+component_path = "transformer/"
+n_layers = 60
+can_shard = true
+safetensors_index_filename = "diffusion_pytorch_model.safetensors.index.json"
+
+[components.storage_size]
+in_bytes = 40860802176
+
+[[components]]
+component_name = "vae"
+component_path = "vae/"
+can_shard = false
+
+[components.storage_size]
+in_bytes = 0

resources/inference_model_cards/mlx-community--DeepSeek-V3.1-4bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/DeepSeek-V3.1-4bit"
+n_layers = 61
+hidden_size = 7168
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "deepseek"
+quantization = "4bit"
+base_model = "DeepSeek V3.1"
+capabilities = ["text", "thinking"]
+
+[storage_size]
+in_bytes = 405874409472

resources/inference_model_cards/mlx-community--DeepSeek-V3.1-8bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/DeepSeek-V3.1-8bit"
+n_layers = 61
+hidden_size = 7168
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "deepseek"
+quantization = "8bit"
+base_model = "DeepSeek V3.1"
+capabilities = ["text", "thinking"]
+
+[storage_size]
+in_bytes = 765577920512

resources/inference_model_cards/mlx-community--GLM-4.5-Air-8bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/GLM-4.5-Air-8bit"
+n_layers = 46
+hidden_size = 4096
+supports_tensor = false
+tasks = ["TextGeneration"]
+family = "glm"
+quantization = "8bit"
+base_model = "GLM 4.5 Air"
+capabilities = ["text", "thinking"]
+
+[storage_size]
+in_bytes = 122406567936

resources/inference_model_cards/mlx-community--GLM-4.5-Air-bf16.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/GLM-4.5-Air-bf16"
+n_layers = 46
+hidden_size = 4096
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "glm"
+quantization = "bf16"
+base_model = "GLM 4.5 Air"
+capabilities = ["text", "thinking"]
+
+[storage_size]
+in_bytes = 229780750336

resources/inference_model_cards/mlx-community--GLM-4.7-4bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/GLM-4.7-4bit"
+n_layers = 91
+hidden_size = 5120
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "glm"
+quantization = "4bit"
+base_model = "GLM 4.7"
+capabilities = ["text", "thinking"]
+
+[storage_size]
+in_bytes = 198556925568

resources/inference_model_cards/mlx-community--GLM-4.7-6bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/GLM-4.7-6bit"
+n_layers = 91
+hidden_size = 5120
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "glm"
+quantization = "6bit"
+base_model = "GLM 4.7"
+capabilities = ["text", "thinking"]
+
+[storage_size]
+in_bytes = 286737579648

resources/inference_model_cards/mlx-community--GLM-4.7-8bit-gs32.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/GLM-4.7-8bit-gs32"
+n_layers = 91
+hidden_size = 5120
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "glm"
+quantization = "8bit"
+base_model = "GLM 4.7"
+capabilities = ["text", "thinking"]
+
+[storage_size]
+in_bytes = 396963397248

resources/inference_model_cards/mlx-community--GLM-4.7-Flash-4bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/GLM-4.7-Flash-4bit"
+n_layers = 47
+hidden_size = 2048
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "glm"
+quantization = "4bit"
+base_model = "GLM 4.7 Flash"
+capabilities = ["text", "thinking"]
+
+[storage_size]
+in_bytes = 19327352832

resources/inference_model_cards/mlx-community--GLM-4.7-Flash-5bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/GLM-4.7-Flash-5bit"
+n_layers = 47
+hidden_size = 2048
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "glm"
+quantization = "5bit"
+base_model = "GLM 4.7 Flash"
+capabilities = ["text", "thinking"]
+
+[storage_size]
+in_bytes = 22548578304

resources/inference_model_cards/mlx-community--GLM-4.7-Flash-6bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/GLM-4.7-Flash-6bit"
+n_layers = 47
+hidden_size = 2048
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "glm"
+quantization = "6bit"
+base_model = "GLM 4.7 Flash"
+capabilities = ["text", "thinking"]
+
+[storage_size]
+in_bytes = 26843545600

resources/inference_model_cards/mlx-community--GLM-4.7-Flash-8bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/GLM-4.7-Flash-8bit"
+n_layers = 47
+hidden_size = 2048
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "glm"
+quantization = "8bit"
+base_model = "GLM 4.7 Flash"
+capabilities = ["text", "thinking"]
+
+[storage_size]
+in_bytes = 34359738368

resources/inference_model_cards/mlx-community--Kimi-K2-Instruct-4bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/Kimi-K2-Instruct-4bit"
+n_layers = 61
+hidden_size = 7168
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "kimi"
+quantization = "4bit"
+base_model = "Kimi K2"
+capabilities = ["text"]
+
+[storage_size]
+in_bytes = 620622774272

resources/inference_model_cards/mlx-community--Kimi-K2-Thinking.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/Kimi-K2-Thinking"
+n_layers = 61
+hidden_size = 7168
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "kimi"
+quantization = ""
+base_model = "Kimi K2"
+capabilities = ["text", "thinking"]
+
+[storage_size]
+in_bytes = 706522120192

resources/inference_model_cards/mlx-community--Kimi-K2.5.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/Kimi-K2.5"
+n_layers = 61
+hidden_size = 7168
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "kimi"
+quantization = ""
+base_model = "Kimi K2.5"
+capabilities = ["text", "thinking"]
+
+[storage_size]
+in_bytes = 662498705408

resources/inference_model_cards/mlx-community--Llama-3.2-1B-Instruct-4bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/Llama-3.2-1B-Instruct-4bit"
+n_layers = 16
+hidden_size = 2048
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "llama"
+quantization = "4bit"
+base_model = "Llama 3.2 1B"
+capabilities = ["text"]
+
+[storage_size]
+in_bytes = 729808896

resources/inference_model_cards/mlx-community--Llama-3.2-3B-Instruct-4bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/Llama-3.2-3B-Instruct-4bit"
+n_layers = 28
+hidden_size = 3072
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "llama"
+quantization = "4bit"
+base_model = "Llama 3.2 3B"
+capabilities = ["text"]
+
+[storage_size]
+in_bytes = 1863319552

resources/inference_model_cards/mlx-community--Llama-3.2-3B-Instruct-8bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/Llama-3.2-3B-Instruct-8bit"
+n_layers = 28
+hidden_size = 3072
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "llama"
+quantization = "8bit"
+base_model = "Llama 3.2 3B"
+capabilities = ["text"]
+
+[storage_size]
+in_bytes = 3501195264

resources/inference_model_cards/mlx-community--Llama-3.3-70B-Instruct-4bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/Llama-3.3-70B-Instruct-4bit"
+n_layers = 80
+hidden_size = 8192
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "llama"
+quantization = "4bit"
+base_model = "Llama 3.3 70B"
+capabilities = ["text"]
+
+[storage_size]
+in_bytes = 40652242944

resources/inference_model_cards/mlx-community--Llama-3.3-70B-Instruct-8bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/Llama-3.3-70B-Instruct-8bit"
+n_layers = 80
+hidden_size = 8192
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "llama"
+quantization = "8bit"
+base_model = "Llama 3.3 70B"
+capabilities = ["text"]
+
+[storage_size]
+in_bytes = 76799803392

resources/inference_model_cards/mlx-community--Meta-Llama-3.1-70B-Instruct-4bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/Meta-Llama-3.1-70B-Instruct-4bit"
+n_layers = 80
+hidden_size = 8192
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "llama"
+quantization = "4bit"
+base_model = "Llama 3.1 70B"
+capabilities = ["text"]
+
+[storage_size]
+in_bytes = 40652242944

resources/inference_model_cards/mlx-community--Meta-Llama-3.1-8B-Instruct-4bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/Meta-Llama-3.1-8B-Instruct-4bit"
+n_layers = 32
+hidden_size = 4096
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "llama"
+quantization = "4bit"
+base_model = "Llama 3.1 8B"
+capabilities = ["text"]
+
+[storage_size]
+in_bytes = 4637851648

resources/inference_model_cards/mlx-community--Meta-Llama-3.1-8B-Instruct-8bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/Meta-Llama-3.1-8B-Instruct-8bit"
+n_layers = 32
+hidden_size = 4096
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "llama"
+quantization = "8bit"
+base_model = "Llama 3.1 8B"
+capabilities = ["text"]
+
+[storage_size]
+in_bytes = 8954839040

resources/inference_model_cards/mlx-community--Meta-Llama-3.1-8B-Instruct-bf16.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/Meta-Llama-3.1-8B-Instruct-bf16"
+n_layers = 32
+hidden_size = 4096
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "llama"
+quantization = "bf16"
+base_model = "Llama 3.1 8B"
+capabilities = ["text"]
+
+[storage_size]
+in_bytes = 16882073600

resources/inference_model_cards/mlx-community--MiniMax-M2.1-3bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/MiniMax-M2.1-3bit"
+n_layers = 61
+hidden_size = 3072
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "minimax"
+quantization = "3bit"
+base_model = "MiniMax M2.1"
+capabilities = ["text", "thinking"]
+
+[storage_size]
+in_bytes = 100086644736

resources/inference_model_cards/mlx-community--MiniMax-M2.1-8bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/MiniMax-M2.1-8bit"
+n_layers = 61
+hidden_size = 3072
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "minimax"
+quantization = "8bit"
+base_model = "MiniMax M2.1"
+capabilities = ["text", "thinking"]
+
+[storage_size]
+in_bytes = 242986745856

resources/inference_model_cards/mlx-community--Qwen3-0.6B-4bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/Qwen3-0.6B-4bit"
+n_layers = 28
+hidden_size = 1024
+supports_tensor = false
+tasks = ["TextGeneration"]
+family = "qwen"
+quantization = "4bit"
+base_model = "Qwen3 0.6B"
+capabilities = ["text", "thinking"]
+
+[storage_size]
+in_bytes = 342884352

resources/inference_model_cards/mlx-community--Qwen3-0.6B-8bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/Qwen3-0.6B-8bit"
+n_layers = 28
+hidden_size = 1024
+supports_tensor = false
+tasks = ["TextGeneration"]
+family = "qwen"
+quantization = "8bit"
+base_model = "Qwen3 0.6B"
+capabilities = ["text", "thinking"]
+
+[storage_size]
+in_bytes = 698351616

resources/inference_model_cards/mlx-community--Qwen3-235B-A22B-Instruct-2507-4bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/Qwen3-235B-A22B-Instruct-2507-4bit"
+n_layers = 94
+hidden_size = 4096
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "qwen"
+quantization = "4bit"
+base_model = "Qwen3 235B"
+capabilities = ["text", "thinking"]
+
+[storage_size]
+in_bytes = 141733920768

resources/inference_model_cards/mlx-community--Qwen3-235B-A22B-Instruct-2507-8bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/Qwen3-235B-A22B-Instruct-2507-8bit"
+n_layers = 94
+hidden_size = 4096
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "qwen"
+quantization = "8bit"
+base_model = "Qwen3 235B"
+capabilities = ["text", "thinking"]
+
+[storage_size]
+in_bytes = 268435456000

resources/inference_model_cards/mlx-community--Qwen3-30B-A3B-4bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/Qwen3-30B-A3B-4bit"
+n_layers = 48
+hidden_size = 2048
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "qwen"
+quantization = "4bit"
+base_model = "Qwen3 30B"
+capabilities = ["text", "thinking"]
+
+[storage_size]
+in_bytes = 17612931072

resources/inference_model_cards/mlx-community--Qwen3-30B-A3B-8bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/Qwen3-30B-A3B-8bit"
+n_layers = 48
+hidden_size = 2048
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "qwen"
+quantization = "8bit"
+base_model = "Qwen3 30B"
+capabilities = ["text", "thinking"]
+
+[storage_size]
+in_bytes = 33279705088

resources/inference_model_cards/mlx-community--Qwen3-Coder-480B-A35B-Instruct-4bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/Qwen3-Coder-480B-A35B-Instruct-4bit"
+n_layers = 62
+hidden_size = 6144
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "qwen"
+quantization = "4bit"
+base_model = "Qwen3 Coder 480B"
+capabilities = ["text", "code"]
+
+[storage_size]
+in_bytes = 289910292480

resources/inference_model_cards/mlx-community--Qwen3-Coder-480B-A35B-Instruct-8bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/Qwen3-Coder-480B-A35B-Instruct-8bit"
+n_layers = 62
+hidden_size = 6144
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "qwen"
+quantization = "8bit"
+base_model = "Qwen3 Coder 480B"
+capabilities = ["text", "code"]
+
+[storage_size]
+in_bytes = 579820584960

resources/inference_model_cards/mlx-community--Qwen3-Next-80B-A3B-Instruct-4bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/Qwen3-Next-80B-A3B-Instruct-4bit"
+n_layers = 48
+hidden_size = 2048
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "qwen"
+quantization = "4bit"
+base_model = "Qwen3 Next 80B"
+capabilities = ["text"]
+
+[storage_size]
+in_bytes = 46976204800

resources/inference_model_cards/mlx-community--Qwen3-Next-80B-A3B-Instruct-8bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/Qwen3-Next-80B-A3B-Instruct-8bit"
+n_layers = 48
+hidden_size = 2048
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "qwen"
+quantization = "8bit"
+base_model = "Qwen3 Next 80B"
+capabilities = ["text"]
+
+[storage_size]
+in_bytes = 88814387200

resources/inference_model_cards/mlx-community--Qwen3-Next-80B-A3B-Thinking-4bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/Qwen3-Next-80B-A3B-Thinking-4bit"
+n_layers = 48
+hidden_size = 2048
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "qwen"
+quantization = "4bit"
+base_model = "Qwen3 Next 80B"
+capabilities = ["text", "thinking"]
+
+[storage_size]
+in_bytes = 47080074240

resources/inference_model_cards/mlx-community--Qwen3-Next-80B-A3B-Thinking-8bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/Qwen3-Next-80B-A3B-Thinking-8bit"
+n_layers = 48
+hidden_size = 2048
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "qwen"
+quantization = "8bit"
+base_model = "Qwen3 Next 80B"
+capabilities = ["text", "thinking"]
+
+[storage_size]
+in_bytes = 88814387200

resources/inference_model_cards/mlx-community--gpt-oss-120b-MXFP4-Q8.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/gpt-oss-120b-MXFP4-Q8"
+n_layers = 36
+hidden_size = 2880
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "gpt-oss"
+quantization = "MXFP4-Q8"
+base_model = "GPT-OSS 120B"
+capabilities = ["text", "thinking"]
+
+[storage_size]
+in_bytes = 70652212224

resources/inference_model_cards/mlx-community--gpt-oss-20b-MXFP4-Q8.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/gpt-oss-20b-MXFP4-Q8"
+n_layers = 24
+hidden_size = 2880
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "gpt-oss"
+quantization = "MXFP4-Q8"
+base_model = "GPT-OSS 20B"
+capabilities = ["text", "thinking"]
+
+[storage_size]
+in_bytes = 12025908224

resources/inference_model_cards/mlx-community--llama-3.3-70b-instruct-fp16.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/llama-3.3-70b-instruct-fp16"
+n_layers = 80
+hidden_size = 8192
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "llama"
+quantization = "fp16"
+base_model = "Llama 3.3 70B"
+capabilities = ["text"]
+
+[storage_size]
+in_bytes = 144383672320

src/exo/download/coordinator.py

@@ -1,4 +1,5 @@
 import asyncio
+import socket
 from dataclasses import dataclass, field
 from typing import Iterator
 
@@ -60,10 +61,37 @@ class DownloadCoordinator:
 
     async def run(self) -> None:
         logger.info("Starting DownloadCoordinator")
+        self._test_internet_connection()
         async with self._tg as tg:
             tg.start_soon(self._command_processor)
             tg.start_soon(self._forward_events)
             tg.start_soon(self._emit_existing_download_progress)
+            tg.start_soon(self._check_internet_connection)
+
+    def _test_internet_connection(self) -> None:
+        try:
+            socket.create_connection(("1.1.1.1", 443), timeout=3).close()
+            self.shard_downloader.set_internet_connection(True)
+        except OSError:
+            self.shard_downloader.set_internet_connection(False)
+        logger.debug(
+            f"Internet connectivity: {self.shard_downloader.internet_connection}"
+        )
+
+    async def _check_internet_connection(self) -> None:
+        first_connection = True
+        while True:
+            await asyncio.sleep(10)
+
+            # Assume that internet connection is set to False on 443 errors.
+            if self.shard_downloader.internet_connection:
+                continue
+
+            self._test_internet_connection()
+
+            if first_connection and self.shard_downloader.internet_connection:
+                first_connection = False
+                self._tg.start_soon(self._emit_existing_download_progress)
 
     def shutdown(self) -> None:
         self._tg.cancel_scope.cancel()
@@ -241,7 +269,7 @@ class DownloadCoordinator:
     async def _emit_existing_download_progress(self) -> None:
         try:
             while True:
-                logger.info(
+                logger.debug(
                     "DownloadCoordinator: Fetching and emitting existing download progress..."
                 )
                 async for (
@@ -274,10 +302,10 @@ class DownloadCoordinator:
                     await self.event_sender.send(
                         NodeDownloadProgress(download_progress=status)
                     )
-                logger.info(
+                logger.debug(
                     "DownloadCoordinator: Done emitting existing download progress."
                 )
-                await anyio.sleep(5 * 60)  # 5 minutes
+                await anyio.sleep(60)
         except Exception as e:
             logger.error(
                 f"DownloadCoordinator: Error emitting existing download progress: {e}"

src/exo/download/download_utils.py

@@ -49,6 +49,10 @@ class HuggingFaceAuthenticationError(Exception):
     """Raised when HuggingFace returns 401/403 for a model download."""
 
 
+class HuggingFaceRateLimitError(Exception):
+    """429 Huggingface code"""
+
+
 async def _build_auth_error_message(status_code: int, model_id: ModelId) -> str:
     token = await get_hf_token()
     if status_code == 401 and token is None:
@@ -154,49 +158,76 @@ async def seed_models(seed_dir: str | Path):
                     logger.error(traceback.format_exc())
 
 
+_fetched_file_lists_this_session: set[str] = set()
+
+
 async def fetch_file_list_with_cache(
-    model_id: ModelId, revision: str = "main", recursive: bool = False
+    model_id: ModelId,
+    revision: str = "main",
+    recursive: bool = False,
+    skip_internet: bool = False,
+    on_connection_lost: Callable[[], None] = lambda: None,
 ) -> list[FileListEntry]:
     target_dir = (await ensure_models_dir()) / "caches" / model_id.normalize()
     await aios.makedirs(target_dir, exist_ok=True)
     cache_file = target_dir / f"{model_id.normalize()}--{revision}--file_list.json"
+    cache_key = f"{model_id.normalize()}--{revision}"
+
+    if cache_key in _fetched_file_lists_this_session and await aios.path.exists(
+        cache_file
+    ):
+        async with aiofiles.open(cache_file, "r") as f:
+            return TypeAdapter(list[FileListEntry]).validate_json(await f.read())
+
+    if skip_internet:
+        if await aios.path.exists(cache_file):
+            async with aiofiles.open(cache_file, "r") as f:
+                return TypeAdapter(list[FileListEntry]).validate_json(await f.read())
+        raise FileNotFoundError(
+            f"No internet connection and no cached file list for {model_id}"
+        )
 
-    # Always try fresh first
     try:
         file_list = await fetch_file_list_with_retry(
-            model_id, revision, recursive=recursive
+            model_id,
+            revision,
+            recursive=recursive,
+            on_connection_lost=on_connection_lost,
         )
-        # Update cache with fresh data
         async with aiofiles.open(cache_file, "w") as f:
             await f.write(
                 TypeAdapter(list[FileListEntry]).dump_json(file_list).decode()
             )
+        _fetched_file_lists_this_session.add(cache_key)
         return file_list
     except Exception as e:
-        # Fetch failed - try cache fallback
         if await aios.path.exists(cache_file):
             logger.warning(
                 f"Failed to fetch file list for {model_id}, using cached data: {e}"
             )
             async with aiofiles.open(cache_file, "r") as f:
                 return TypeAdapter(list[FileListEntry]).validate_json(await f.read())
-        # No cache available, propagate the error
-        raise
+        raise FileNotFoundError(f"Failed to fetch file list for {model_id}: {e}") from e
 
 
 async def fetch_file_list_with_retry(
-    model_id: ModelId, revision: str = "main", path: str = "", recursive: bool = False
+    model_id: ModelId,
+    revision: str = "main",
+    path: str = "",
+    recursive: bool = False,
+    on_connection_lost: Callable[[], None] = lambda: None,
 ) -> list[FileListEntry]:
-    n_attempts = 30
+    n_attempts = 3
     for attempt in range(n_attempts):
         try:
             return await _fetch_file_list(model_id, revision, path, recursive)
         except HuggingFaceAuthenticationError:
             raise
         except Exception as e:
+            on_connection_lost()
             if attempt == n_attempts - 1:
                 raise e
-            await asyncio.sleep(min(8, 0.1 * float(2.0 ** int(attempt))))
+            await asyncio.sleep(2.0**attempt)
     raise Exception(
         f"Failed to fetch file list for {model_id=} {revision=} {path=} {recursive=}"
     )
@@ -216,7 +247,11 @@ async def _fetch_file_list(
         if response.status in [401, 403]:
             msg = await _build_auth_error_message(response.status, model_id)
             raise HuggingFaceAuthenticationError(msg)
-        if response.status == 200:
+        elif response.status == 429:
+            raise HuggingFaceRateLimitError(
+                f"Couldn't download {model_id} because of HuggingFace rate limit."
+            )
+        elif response.status == 200:
             data_json = await response.text()
             data = TypeAdapter(list[FileListEntry]).validate_json(data_json)
             files: list[FileListEntry] = []
@@ -249,7 +284,7 @@ def create_http_session(
     else:
         total_timeout = 1800
         connect_timeout = 60
-        sock_read_timeout = 1800
+        sock_read_timeout = 60
         sock_connect_timeout = 60
 
     ssl_context = ssl.create_default_context(
@@ -324,8 +359,9 @@ async def download_file_with_retry(
     path: str,
     target_dir: Path,
     on_progress: Callable[[int, int, bool], None] = lambda _, __, ___: None,
+    on_connection_lost: Callable[[], None] = lambda: None,
 ) -> Path:
-    n_attempts = 30
+    n_attempts = 3
     for attempt in range(n_attempts):
         try:
             return await _download_file(
@@ -333,14 +369,19 @@ async def download_file_with_retry(
             )
         except HuggingFaceAuthenticationError:
             raise
-        except Exception as e:
-            if isinstance(e, FileNotFoundError) or attempt == n_attempts - 1:
+        except HuggingFaceRateLimitError as e:
+            if attempt == n_attempts - 1:
                 raise e
             logger.error(
                 f"Download error on attempt {attempt}/{n_attempts} for {model_id=} {revision=} {path=} {target_dir=}"
             )
             logger.error(traceback.format_exc())
-            await asyncio.sleep(min(8, 0.1 * (2.0**attempt)))
+            await asyncio.sleep(2.0**attempt)
+        except Exception as e:
+            on_connection_lost()
+            if attempt == n_attempts - 1:
+                raise e
+            break
     raise Exception(
         f"Failed to download file {model_id=} {revision=} {path=} {target_dir=}"
     )
@@ -542,7 +583,9 @@ async def download_shard(
     on_progress: Callable[[ShardMetadata, RepoDownloadProgress], Awaitable[None]],
     max_parallel_downloads: int = 8,
     skip_download: bool = False,
+    skip_internet: bool = False,
     allow_patterns: list[str] | None = None,
+    on_connection_lost: Callable[[], None] = lambda: None,
 ) -> tuple[Path, RepoDownloadProgress]:
     if not skip_download:
         logger.debug(f"Downloading {shard.model_card.model_id=}")
@@ -562,7 +605,11 @@ async def download_shard(
 
     all_start_time = time.time()
     file_list = await fetch_file_list_with_cache(
-        shard.model_card.model_id, revision, recursive=True
+        shard.model_card.model_id,
+        revision,
+        recursive=True,
+        skip_internet=skip_internet,
+        on_connection_lost=on_connection_lost,
     )
     filtered_file_list = list(
         filter_repo_objects(
@@ -672,6 +719,7 @@ async def download_shard(
                 lambda curr_bytes, total_bytes, is_renamed: schedule_progress(
                     file, curr_bytes, total_bytes, is_renamed
                 ),
+                on_connection_lost=on_connection_lost,
             )
 
     if not skip_download:

src/exo/download/impl_shard_downloader.py

@@ -1,4 +1,5 @@
 import asyncio
+from asyncio import create_task
 from collections.abc import Awaitable
 from pathlib import Path
 from typing import AsyncIterator, Callable
@@ -7,7 +8,7 @@ from loguru import logger
 
 from exo.download.download_utils import RepoDownloadProgress, download_shard
 from exo.download.shard_downloader import ShardDownloader
-from exo.shared.models.model_cards import MODEL_CARDS, ModelCard, ModelId
+from exo.shared.models.model_cards import ModelCard, ModelId, get_model_cards
 from exo.shared.types.worker.shards import (
     PipelineShardMetadata,
     ShardMetadata,
@@ -49,6 +50,10 @@ class SingletonShardDownloader(ShardDownloader):
         self.shard_downloader = shard_downloader
         self.active_downloads: dict[ShardMetadata, asyncio.Task[Path]] = {}
 
+    def set_internet_connection(self, value: bool) -> None:
+        self.internet_connection = value
+        self.shard_downloader.set_internet_connection(value)
+
     def on_progress(
         self,
         callback: Callable[[ShardMetadata, RepoDownloadProgress], Awaitable[None]],
@@ -85,6 +90,10 @@ class CachedShardDownloader(ShardDownloader):
         self.shard_downloader = shard_downloader
         self.cache: dict[tuple[str, ShardMetadata], Path] = {}
 
+    def set_internet_connection(self, value: bool) -> None:
+        self.internet_connection = value
+        self.shard_downloader.set_internet_connection(value)
+
     def on_progress(
         self,
         callback: Callable[[ShardMetadata, RepoDownloadProgress], Awaitable[None]],
@@ -142,6 +151,8 @@ class ResumableShardDownloader(ShardDownloader):
             self.on_progress_wrapper,
             max_parallel_downloads=self.max_parallel_downloads,
             allow_patterns=allow_patterns,
+            skip_internet=not self.internet_connection,
+            on_connection_lost=lambda: self.set_internet_connection(False),
         )
         return target_dir
 
@@ -154,13 +165,24 @@ class ResumableShardDownloader(ShardDownloader):
             """Helper coroutine that builds the shard for a model and gets its download status."""
             shard = await build_full_shard(model_id)
             return await download_shard(
-                shard, self.on_progress_wrapper, skip_download=True
+                shard,
+                self.on_progress_wrapper,
+                skip_download=True,
+                skip_internet=not self.internet_connection,
+                on_connection_lost=lambda: self.set_internet_connection(False),
             )
 
-        # Kick off download status coroutines concurrently
+        semaphore = asyncio.Semaphore(self.max_parallel_downloads)
+
+        async def download_with_semaphore(
+            model_card: ModelCard,
+        ) -> tuple[Path, RepoDownloadProgress]:
+            async with semaphore:
+                return await _status_for_model(model_card.model_id)
+
         tasks = [
-            asyncio.create_task(_status_for_model(model_card.model_id))
-            for model_card in MODEL_CARDS.values()
+            create_task(download_with_semaphore(model_card))
+            for model_card in await get_model_cards()
         ]
 
         for task in asyncio.as_completed(tasks):

src/exo/download/shard_downloader.py

@@ -16,6 +16,11 @@ from exo.shared.types.worker.shards import (
 
 # TODO: the PipelineShardMetadata getting reinstantiated is a bit messy. Should this be a classmethod?
 class ShardDownloader(ABC):
+    internet_connection: bool = False
+
+    def set_internet_connection(self, value: bool) -> None:
+        self.internet_connection = value
+
     @abstractmethod
     async def ensure_shard(
         self, shard: ShardMetadata, config_only: bool = False

src/exo/main.py

@@ -90,7 +90,6 @@ class Node:
             worker = Worker(
                 node_id,
                 session_id,
-                connection_message_receiver=router.receiver(topics.CONNECTION_MESSAGES),
                 global_event_receiver=router.receiver(topics.GLOBAL_EVENTS),
                 local_event_sender=router.sender(topics.LOCAL_EVENTS),
                 command_sender=router.sender(topics.COMMANDS),
@@ -227,9 +226,6 @@ class Node:
                         self.worker = Worker(
                             self.node_id,
                             result.session_id,
-                            connection_message_receiver=self.router.receiver(
-                                topics.CONNECTION_MESSAGES
-                            ),
                             global_event_receiver=self.router.receiver(
                                 topics.GLOBAL_EVENTS
                             ),

src/exo/master/adapters/__init__.py

@@ -0,0 +1 @@
+"""API adapters for different API formats (Claude, OpenAI Responses, etc.)."""

src/exo/master/adapters/chat_completions.py

@@ -0,0 +1,214 @@
+"""OpenAI Chat Completions API adapter for converting requests/responses."""
+
+import time
+from collections.abc import AsyncGenerator
+from typing import Any
+from uuid import uuid4
+
+from exo.shared.types.api import (
+    ChatCompletionChoice,
+    ChatCompletionMessage,
+    ChatCompletionMessageText,
+    ChatCompletionRequest,
+    ChatCompletionResponse,
+    ErrorInfo,
+    ErrorResponse,
+    FinishReason,
+    StreamingChoiceResponse,
+    ToolCall,
+)
+from exo.shared.types.chunks import ErrorChunk, TokenChunk, ToolCallChunk
+from exo.shared.types.common import CommandId
+from exo.shared.types.text_generation import InputMessage, TextGenerationTaskParams
+
+
+def chat_request_to_text_generation(
+    request: ChatCompletionRequest,
+) -> TextGenerationTaskParams:
+    instructions: str | None = None
+    input_messages: list[InputMessage] = []
+    chat_template_messages: list[dict[str, Any]] = []
+
+    for msg in request.messages:
+        # Normalize content to string
+        content: str
+        if msg.content is None:
+            content = ""
+        elif isinstance(msg.content, str):
+            content = msg.content
+        elif isinstance(msg.content, ChatCompletionMessageText):
+            content = msg.content.text
+        else:
+            # List of ChatCompletionMessageText
+            content = "\n".join(item.text for item in msg.content)
+
+        # Extract system message as instructions
+        if msg.role == "system":
+            if instructions is None:
+                instructions = content
+            else:
+                # Append additional system messages
+                instructions = f"{instructions}\n{content}"
+            chat_template_messages.append({"role": "system", "content": content})
+        else:
+            # Skip messages with no meaningful content
+            if msg.content is None and msg.thinking is None and msg.tool_calls is None:
+                continue
+
+            if msg.role in ("user", "assistant", "developer"):
+                input_messages.append(InputMessage(role=msg.role, content=content))
+
+            # Build full message dict for chat template (preserves tool_calls etc.)
+            # Normalize content for model_dump
+            msg_copy = msg.model_copy(update={"content": content})
+            dumped: dict[str, Any] = msg_copy.model_dump(exclude_none=True)
+            chat_template_messages.append(dumped)
+
+    return TextGenerationTaskParams(
+        model=request.model,
+        input=input_messages
+        if input_messages
+        else [InputMessage(role="user", content="")],
+        instructions=instructions,
+        max_output_tokens=request.max_tokens,
+        temperature=request.temperature,
+        top_p=request.top_p,
+        top_k=request.top_k,
+        stop=request.stop,
+        seed=request.seed,
+        stream=request.stream,
+        tools=request.tools,
+        chat_template_messages=chat_template_messages
+        if chat_template_messages
+        else None,
+    )
+
+
+def chunk_to_response(
+    chunk: TokenChunk, command_id: CommandId
+) -> ChatCompletionResponse:
+    """Convert a TokenChunk to a streaming ChatCompletionResponse."""
+    return ChatCompletionResponse(
+        id=command_id,
+        created=int(time.time()),
+        model=chunk.model,
+        choices=[
+            StreamingChoiceResponse(
+                index=0,
+                delta=ChatCompletionMessage(role="assistant", content=chunk.text),
+                finish_reason=chunk.finish_reason,
+            )
+        ],
+    )
+
+
+async def generate_chat_stream(
+    command_id: CommandId,
+    chunk_stream: AsyncGenerator[ErrorChunk | ToolCallChunk | TokenChunk, None],
+) -> AsyncGenerator[str, None]:
+    """Generate Chat Completions API streaming events from chunks."""
+    async for chunk in chunk_stream:
+        if isinstance(chunk, ErrorChunk):
+            error_response = ErrorResponse(
+                error=ErrorInfo(
+                    message=chunk.error_message or "Internal server error",
+                    type="InternalServerError",
+                    code=500,
+                )
+            )
+            yield f"data: {error_response.model_dump_json()}\n\n"
+            yield "data: [DONE]\n\n"
+            return
+
+        if isinstance(chunk, ToolCallChunk):
+            tool_call_deltas = [
+                ToolCall(
+                    id=str(uuid4()),
+                    index=i,
+                    function=tool,
+                )
+                for i, tool in enumerate(chunk.tool_calls)
+            ]
+            tool_response = ChatCompletionResponse(
+                id=command_id,
+                created=int(time.time()),
+                model=chunk.model,
+                choices=[
+                    StreamingChoiceResponse(
+                        index=0,
+                        delta=ChatCompletionMessage(
+                            role="assistant",
+                            tool_calls=tool_call_deltas,
+                        ),
+                        finish_reason="tool_calls",
+                    )
+                ],
+            )
+            yield f"data: {tool_response.model_dump_json()}\n\n"
+            yield "data: [DONE]\n\n"
+            return
+
+        chunk_response = chunk_to_response(chunk, command_id)
+        yield f"data: {chunk_response.model_dump_json()}\n\n"
+
+        if chunk.finish_reason is not None:
+            yield "data: [DONE]\n\n"
+
+
+async def collect_chat_response(
+    command_id: CommandId,
+    chunk_stream: AsyncGenerator[ErrorChunk | ToolCallChunk | TokenChunk, None],
+) -> ChatCompletionResponse:
+    """Collect all token chunks and return a single ChatCompletionResponse."""
+    text_parts: list[str] = []
+    tool_calls: list[ToolCall] = []
+    model: str | None = None
+    finish_reason: FinishReason | None = None
+    error_message: str | None = None
+
+    async for chunk in chunk_stream:
+        if isinstance(chunk, ErrorChunk):
+            error_message = chunk.error_message or "Internal server error"
+            break
+
+        if model is None:
+            model = chunk.model
+
+        if isinstance(chunk, TokenChunk):
+            text_parts.append(chunk.text)
+
+        if isinstance(chunk, ToolCallChunk):
+            tool_calls.extend(
+                ToolCall(
+                    id=str(uuid4()),
+                    index=i,
+                    function=tool,
+                )
+                for i, tool in enumerate(chunk.tool_calls)
+            )
+
+        if chunk.finish_reason is not None:
+            finish_reason = chunk.finish_reason
+
+    if error_message is not None:
+        raise ValueError(error_message)
+
+    combined_text = "".join(text_parts)
+    assert model is not None
+
+    return ChatCompletionResponse(
+        id=command_id,
+        created=int(time.time()),
+        model=model,
+        choices=[
+            ChatCompletionChoice(
+                index=0,
+                message=ChatCompletionMessage(
+                    role="assistant",
+                    content=combined_text,
+                    tool_calls=tool_calls if tool_calls else None,
+                ),
+                finish_reason=finish_reason,
+            )
+        ],
+    )

src/exo/master/adapters/claude.py

@@ -0,0 +1,323 @@
+"""Claude Messages API adapter for converting requests/responses."""
+
+import json
+from collections.abc import AsyncGenerator
+from typing import Any
+from uuid import uuid4
+
+from exo.shared.types.api import FinishReason
+from exo.shared.types.chunks import ErrorChunk, TokenChunk, ToolCallChunk
+from exo.shared.types.claude_api import (
+    ClaudeContentBlock,
+    ClaudeContentBlockDeltaEvent,
+    ClaudeContentBlockStartEvent,
+    ClaudeContentBlockStopEvent,
+    ClaudeInputJsonDelta,
+    ClaudeMessageDelta,
+    ClaudeMessageDeltaEvent,
+    ClaudeMessageDeltaUsage,
+    ClaudeMessagesRequest,
+    ClaudeMessagesResponse,
+    ClaudeMessageStart,
+    ClaudeMessageStartEvent,
+    ClaudeMessageStopEvent,
+    ClaudeStopReason,
+    ClaudeTextBlock,
+    ClaudeTextDelta,
+    ClaudeToolResultBlock,
+    ClaudeToolUseBlock,
+    ClaudeUsage,
+)
+from exo.shared.types.common import CommandId
+from exo.shared.types.text_generation import InputMessage, TextGenerationTaskParams
+
+
+def finish_reason_to_claude_stop_reason(
+    finish_reason: FinishReason | None,
+) -> ClaudeStopReason | None:
+    """Map OpenAI finish_reason to Claude stop_reason."""
+    if finish_reason is None:
+        return None
+    mapping: dict[FinishReason, ClaudeStopReason] = {
+        "stop": "end_turn",
+        "length": "max_tokens",
+        "tool_calls": "tool_use",
+        "content_filter": "end_turn",
+        "function_call": "tool_use",
+    }
+    return mapping.get(finish_reason, "end_turn")
+
+
+def _extract_tool_result_text(block: ClaudeToolResultBlock) -> str:
+    """Extract plain text from a tool_result content field."""
+    if block.content is None:
+        return ""
+    if isinstance(block.content, str):
+        return block.content
+    return "".join(sub_block.text for sub_block in block.content)
+
+
+def claude_request_to_text_generation(
+    request: ClaudeMessagesRequest,
+) -> TextGenerationTaskParams:
+    # Handle system message
+    instructions: str | None = None
+    chat_template_messages: list[dict[str, Any]] = []
+
+    if request.system:
+        if isinstance(request.system, str):
+            instructions = request.system
+        else:
+            instructions = "".join(block.text for block in request.system)
+        chat_template_messages.append({"role": "system", "content": instructions})
+
+    # Convert messages to input
+    input_messages: list[InputMessage] = []
+    for msg in request.messages:
+        if isinstance(msg.content, str):
+            input_messages.append(InputMessage(role=msg.role, content=msg.content))
+            chat_template_messages.append({"role": msg.role, "content": msg.content})
+            continue
+
+        # Process structured content blocks
+        text_parts: list[str] = []
+        tool_calls: list[dict[str, Any]] = []
+        tool_results: list[ClaudeToolResultBlock] = []
+
+        for block in msg.content:
+            if isinstance(block, ClaudeTextBlock):
+                text_parts.append(block.text)
+            elif isinstance(block, ClaudeToolUseBlock):
+                tool_calls.append(
+                    {
+                        "id": block.id,
+                        "type": "function",
+                        "function": {
+                            "name": block.name,
+                            "arguments": json.dumps(block.input),
+                        },
+                    }
+                )
+            elif isinstance(block, ClaudeToolResultBlock):
+                tool_results.append(block)
+
+        content = "".join(text_parts)
+
+        # Build InputMessage from text content
+        if msg.role in ("user", "assistant"):
+            input_messages.append(InputMessage(role=msg.role, content=content))
+
+        # Build chat_template_messages preserving tool structure
+        if tool_calls:
+            chat_template_messages.append(
+                {"role": "assistant", "content": content, "tool_calls": tool_calls}
+            )
+        elif tool_results:
+            for tr in tool_results:
+                chat_template_messages.append(
+                    {
+                        "role": "tool",
+                        "tool_call_id": tr.tool_use_id,
+                        "content": _extract_tool_result_text(tr),
+                    }
+                )
+        else:
+            chat_template_messages.append({"role": msg.role, "content": content})
+
+    # Convert Claude tool definitions to OpenAI-style function tools
+    tools: list[dict[str, Any]] | None = None
+    if request.tools:
+        tools = [
+            {
+                "type": "function",
+                "function": {
+                    "name": tool.name,
+                    "description": tool.description or "",
+                    "parameters": tool.input_schema,
+                },
+            }
+            for tool in request.tools
+        ]
+
+    return TextGenerationTaskParams(
+        model=request.model,
+        input=input_messages
+        if input_messages
+        else [InputMessage(role="user", content="")],
+        instructions=instructions,
+        max_output_tokens=request.max_tokens,
+        temperature=request.temperature,
+        top_p=request.top_p,
+        top_k=request.top_k,
+        stop=request.stop_sequences,
+        stream=request.stream,
+        tools=tools,
+        chat_template_messages=chat_template_messages
+        if chat_template_messages
+        else None,
+    )
+
+
+async def collect_claude_response(
+    command_id: CommandId,
+    model: str,
+    chunk_stream: AsyncGenerator[ErrorChunk | ToolCallChunk | TokenChunk, None],
+) -> ClaudeMessagesResponse:
+    """Collect all token chunks and return a single ClaudeMessagesResponse."""
+    text_parts: list[str] = []
+    tool_use_blocks: list[ClaudeToolUseBlock] = []
+    stop_reason: ClaudeStopReason | None = None
+    last_stats = None
+    error_message: str | None = None
+
+    async for chunk in chunk_stream:
+        if isinstance(chunk, ErrorChunk):
+            error_message = chunk.error_message or "Internal server error"
+            break
+
+        if isinstance(chunk, ToolCallChunk):
+            for tool in chunk.tool_calls:
+                tool_use_blocks.append(
+                    ClaudeToolUseBlock(
+                        id=f"toolu_{uuid4().hex[:24]}",
+                        name=tool.name,
+                        input=json.loads(tool.arguments),  # pyright: ignore[reportAny]
+                    )
+                )
+            last_stats = chunk.stats or last_stats
+            stop_reason = "tool_use"
+            continue
+
+        text_parts.append(chunk.text)
+        last_stats = chunk.stats or last_stats
+
+        if chunk.finish_reason is not None:
+            stop_reason = finish_reason_to_claude_stop_reason(chunk.finish_reason)
+
+    if error_message is not None:
+        raise ValueError(error_message)
+
+    combined_text = "".join(text_parts)
+
+    # Build content blocks
+    content: list[ClaudeContentBlock] = []
+    if combined_text:
+        content.append(ClaudeTextBlock(text=combined_text))
+    content.extend(tool_use_blocks)
+
+    # If no content at all, include empty text block
+    if not content:
+        content.append(ClaudeTextBlock(text=""))
+
+    # Use actual usage data from stats if available
+    input_tokens = last_stats.prompt_tokens if last_stats else 0
+    output_tokens = last_stats.generation_tokens if last_stats else 0
+
+    return ClaudeMessagesResponse(
+        id=f"msg_{command_id}",
+        model=model,
+        content=content,
+        stop_reason=stop_reason,
+        usage=ClaudeUsage(
+            input_tokens=input_tokens,
+            output_tokens=output_tokens,
+        ),
+    )
+
+
+async def generate_claude_stream(
+    command_id: CommandId,
+    model: str,
+    chunk_stream: AsyncGenerator[ErrorChunk | ToolCallChunk | TokenChunk, None],
+) -> AsyncGenerator[str, None]:
+    """Generate Claude Messages API streaming events from TokenChunks."""
+    # Initial message_start event
+    initial_message = ClaudeMessageStart(
+        id=f"msg_{command_id}",
+        model=model,
+        content=[],
+        stop_reason=None,
+        usage=ClaudeUsage(input_tokens=0, output_tokens=0),
+    )
+    start_event = ClaudeMessageStartEvent(message=initial_message)
+    yield f"event: message_start\ndata: {start_event.model_dump_json()}\n\n"
+
+    # content_block_start for text block at index 0
+    block_start = ClaudeContentBlockStartEvent(
+        index=0, content_block=ClaudeTextBlock(text="")
+    )
+    yield f"event: content_block_start\ndata: {block_start.model_dump_json()}\n\n"
+
+    output_tokens = 0
+    stop_reason: ClaudeStopReason | None = None
+    last_stats = None
+    next_block_index = 1  # text block is 0, tool blocks start at 1
+
+    async for chunk in chunk_stream:
+        if isinstance(chunk, ErrorChunk):
+            # Close text block and bail
+            break
+
+        if isinstance(chunk, ToolCallChunk):
+            last_stats = chunk.stats or last_stats
+            stop_reason = "tool_use"
+
+            # Emit tool_use content blocks
+            for tool in chunk.tool_calls:
+                tool_id = f"toolu_{uuid4().hex[:24]}"
+                tool_input_json = tool.arguments
+
+                # content_block_start for tool_use
+                tool_block_start = ClaudeContentBlockStartEvent(
+                    index=next_block_index,
+                    content_block=ClaudeToolUseBlock(
+                        id=tool_id, name=tool.name, input={}
+                    ),
+                )
+                yield f"event: content_block_start\ndata: {tool_block_start.model_dump_json()}\n\n"
+
+                # content_block_delta with input_json_delta
+                tool_delta_event = ClaudeContentBlockDeltaEvent(
+                    index=next_block_index,
+                    delta=ClaudeInputJsonDelta(partial_json=tool_input_json),
+                )
+                yield f"event: content_block_delta\ndata: {tool_delta_event.model_dump_json()}\n\n"
+
+                # content_block_stop
+                tool_block_stop = ClaudeContentBlockStopEvent(index=next_block_index)
+                yield f"event: content_block_stop\ndata: {tool_block_stop.model_dump_json()}\n\n"
+
+                next_block_index += 1
+            continue
+
+        output_tokens += 1  # Count each chunk as one token
+        last_stats = chunk.stats or last_stats
+
+        # content_block_delta
+        delta_event = ClaudeContentBlockDeltaEvent(
+            index=0,
+            delta=ClaudeTextDelta(text=chunk.text),
+        )
+        yield f"event: content_block_delta\ndata: {delta_event.model_dump_json()}\n\n"
+
+        if chunk.finish_reason is not None:
+            stop_reason = finish_reason_to_claude_stop_reason(chunk.finish_reason)
+
+    # Use actual token count from stats if available
+    if last_stats is not None:
+        output_tokens = last_stats.generation_tokens
+
+    # content_block_stop for text block
+    block_stop = ClaudeContentBlockStopEvent(index=0)
+    yield f"event: content_block_stop\ndata: {block_stop.model_dump_json()}\n\n"
+
+    # message_delta
+    message_delta = ClaudeMessageDeltaEvent(
+        delta=ClaudeMessageDelta(stop_reason=stop_reason),
+        usage=ClaudeMessageDeltaUsage(output_tokens=output_tokens),
+    )
+    yield f"event: message_delta\ndata: {message_delta.model_dump_json()}\n\n"
+
+    # message_stop
+    message_stop = ClaudeMessageStopEvent()
+    yield f"event: message_stop\ndata: {message_stop.model_dump_json()}\n\n"

src/exo/master/adapters/responses.py

@@ -0,0 +1,373 @@
+"""OpenAI Responses API adapter for converting requests/responses."""
+
+from collections.abc import AsyncGenerator
+from itertools import count
+from typing import Any
+from uuid import uuid4
+
+from exo.shared.types.chunks import ErrorChunk, TokenChunk, ToolCallChunk
+from exo.shared.types.common import CommandId
+from exo.shared.types.openai_responses import (
+    FunctionCallInputItem,
+    ResponseCompletedEvent,
+    ResponseContentPart,
+    ResponseContentPartAddedEvent,
+    ResponseContentPartDoneEvent,
+    ResponseCreatedEvent,
+    ResponseFunctionCallArgumentsDeltaEvent,
+    ResponseFunctionCallArgumentsDoneEvent,
+    ResponseFunctionCallItem,
+    ResponseInProgressEvent,
+    ResponseInputMessage,
+    ResponseItem,
+    ResponseMessageItem,
+    ResponseOutputItemAddedEvent,
+    ResponseOutputItemDoneEvent,
+    ResponseOutputText,
+    ResponsesRequest,
+    ResponsesResponse,
+    ResponseTextDeltaEvent,
+    ResponseTextDoneEvent,
+    ResponseUsage,
+)
+from exo.shared.types.text_generation import InputMessage, TextGenerationTaskParams
+
+
+def _extract_content(content: str | list[ResponseContentPart]) -> str:
+    """Extract plain text from a content field that may be a string or list of parts."""
+    if isinstance(content, str):
+        return content
+    return "".join(part.text for part in content)
+
+
+def responses_request_to_text_generation(
+    request: ResponsesRequest,
+) -> TextGenerationTaskParams:
+    input_value: list[InputMessage]
+    built_chat_template: list[dict[str, Any]] | None = None
+    if isinstance(request.input, str):
+        input_value = [InputMessage(role="user", content=request.input)]
+    else:
+        input_messages: list[InputMessage] = []
+        chat_template_messages: list[dict[str, Any]] = []
+
+        if request.instructions is not None:
+            chat_template_messages.append(
+                {"role": "system", "content": request.instructions}
+            )
+
+        for item in request.input:
+            if isinstance(item, ResponseInputMessage):
+                content = _extract_content(item.content)
+                if item.role in ("user", "assistant", "developer"):
+                    input_messages.append(InputMessage(role=item.role, content=content))
+                if item.role == "system":
+                    chat_template_messages.append(
+                        {"role": "system", "content": content}
+                    )
+                else:
+                    chat_template_messages.append(
+                        {"role": item.role, "content": content}
+                    )
+            elif isinstance(item, FunctionCallInputItem):
+                chat_template_messages.append(
+                    {
+                        "role": "assistant",
+                        "content": "",
+                        "tool_calls": [
+                            {
+                                "id": item.call_id,
+                                "type": "function",
+                                "function": {
+                                    "name": item.name,
+                                    "arguments": item.arguments,
+                                },
+                            }
+                        ],
+                    }
+                )
+            else:
+                chat_template_messages.append(
+                    {
+                        "role": "tool",
+                        "tool_call_id": item.call_id,
+                        "content": item.output,
+                    }
+                )
+
+        input_value = (
+            input_messages
+            if input_messages
+            else [InputMessage(role="user", content="")]
+        )
+        built_chat_template = chat_template_messages if chat_template_messages else None
+
+    return TextGenerationTaskParams(
+        model=request.model,
+        input=input_value,
+        instructions=request.instructions,
+        max_output_tokens=request.max_output_tokens,
+        temperature=request.temperature,
+        top_p=request.top_p,
+        stream=request.stream,
+        tools=request.tools,
+        top_k=request.top_k,
+        stop=request.stop,
+        seed=request.seed,
+        chat_template_messages=built_chat_template or request.chat_template_messages,
+    )
+
+
+async def collect_responses_response(
+    command_id: CommandId,
+    model: str,
+    chunk_stream: AsyncGenerator[ErrorChunk | ToolCallChunk | TokenChunk, None],
+) -> ResponsesResponse:
+    """Collect all token chunks and return a single ResponsesResponse."""
+    response_id = f"resp_{command_id}"
+    item_id = f"item_{command_id}"
+    accumulated_text = ""
+    function_call_items: list[ResponseFunctionCallItem] = []
+    last_stats = None
+    error_message: str | None = None
+
+    async for chunk in chunk_stream:
+        if isinstance(chunk, ErrorChunk):
+            error_message = chunk.error_message or "Internal server error"
+            break
+
+        if isinstance(chunk, ToolCallChunk):
+            for tool in chunk.tool_calls:
+                function_call_items.append(
+                    ResponseFunctionCallItem(
+                        id=f"fc_{uuid4().hex[:24]}",
+                        call_id=f"call_{uuid4().hex[:24]}",
+                        name=tool.name,
+                        arguments=tool.arguments,
+                    )
+                )
+            last_stats = chunk.stats or last_stats
+            continue
+
+        accumulated_text += chunk.text
+        last_stats = chunk.stats or last_stats
+
+    if error_message is not None:
+        raise ValueError(error_message)
+
+    # Create usage from stats if available
+    usage = None
+    if last_stats is not None:
+        usage = ResponseUsage(
+            input_tokens=last_stats.prompt_tokens,
+            output_tokens=last_stats.generation_tokens,
+            total_tokens=last_stats.prompt_tokens + last_stats.generation_tokens,
+        )
+
+    output: list[ResponseItem] = [
+        ResponseMessageItem(
+            id=item_id,
+            content=[ResponseOutputText(text=accumulated_text)],
+            status="completed",
+        )
+    ]
+    output.extend(function_call_items)
+
+    return ResponsesResponse(
+        id=response_id,
+        model=model,
+        status="completed",
+        output=output,
+        output_text=accumulated_text,
+        usage=usage,
+    )
+
+
+async def generate_responses_stream(
+    command_id: CommandId,
+    model: str,
+    chunk_stream: AsyncGenerator[ErrorChunk | ToolCallChunk | TokenChunk, None],
+) -> AsyncGenerator[str, None]:
+    """Generate OpenAI Responses API streaming events from TokenChunks."""
+    response_id = f"resp_{command_id}"
+    item_id = f"item_{command_id}"
+    seq = count(1)
+
+    # response.created
+    initial_response = ResponsesResponse(
+        id=response_id,
+        model=model,
+        status="in_progress",
+        output=[],
+        output_text="",
+    )
+    created_event = ResponseCreatedEvent(
+        sequence_number=next(seq), response=initial_response
+    )
+    yield f"event: response.created\ndata: {created_event.model_dump_json()}\n\n"
+
+    # response.in_progress
+    in_progress_event = ResponseInProgressEvent(
+        sequence_number=next(seq), response=initial_response
+    )
+    yield f"event: response.in_progress\ndata: {in_progress_event.model_dump_json()}\n\n"
+
+    # response.output_item.added
+    initial_item = ResponseMessageItem(
+        id=item_id,
+        content=[ResponseOutputText(text="")],
+        status="in_progress",
+    )
+    item_added = ResponseOutputItemAddedEvent(
+        sequence_number=next(seq), output_index=0, item=initial_item
+    )
+    yield f"event: response.output_item.added\ndata: {item_added.model_dump_json()}\n\n"
+
+    # response.content_part.added
+    initial_part = ResponseOutputText(text="")
+    part_added = ResponseContentPartAddedEvent(
+        sequence_number=next(seq),
+        item_id=item_id,
+        output_index=0,
+        content_index=0,
+        part=initial_part,
+    )
+    yield f"event: response.content_part.added\ndata: {part_added.model_dump_json()}\n\n"
+
+    accumulated_text = ""
+    function_call_items: list[ResponseFunctionCallItem] = []
+    last_stats = None
+    next_output_index = 1  # message item is at 0
+
+    async for chunk in chunk_stream:
+        if isinstance(chunk, ErrorChunk):
+            break
+
+        if isinstance(chunk, ToolCallChunk):
+            last_stats = chunk.stats or last_stats
+            for tool in chunk.tool_calls:
+                fc_id = f"fc_{uuid4().hex[:24]}"
+                call_id = f"call_{uuid4().hex[:24]}"
+
+                # response.output_item.added for function_call
+                fc_item = ResponseFunctionCallItem(
+                    id=fc_id,
+                    call_id=call_id,
+                    name=tool.name,
+                    arguments="",
+                    status="in_progress",
+                )
+                fc_added = ResponseOutputItemAddedEvent(
+                    sequence_number=next(seq),
+                    output_index=next_output_index,
+                    item=fc_item,
+                )
+                yield f"event: response.output_item.added\ndata: {fc_added.model_dump_json()}\n\n"
+
+                # response.function_call_arguments.delta
+                args_delta = ResponseFunctionCallArgumentsDeltaEvent(
+                    sequence_number=next(seq),
+                    item_id=fc_id,
+                    output_index=next_output_index,
+                    delta=tool.arguments,
+                )
+                yield f"event: response.function_call_arguments.delta\ndata: {args_delta.model_dump_json()}\n\n"
+
+                # response.function_call_arguments.done
+                args_done = ResponseFunctionCallArgumentsDoneEvent(
+                    sequence_number=next(seq),
+                    item_id=fc_id,
+                    output_index=next_output_index,
+                    name=tool.name,
+                    arguments=tool.arguments,
+                )
+                yield f"event: response.function_call_arguments.done\ndata: {args_done.model_dump_json()}\n\n"
+
+                # response.output_item.done
+                fc_done_item = ResponseFunctionCallItem(
+                    id=fc_id,
+                    call_id=call_id,
+                    name=tool.name,
+                    arguments=tool.arguments,
+                    status="completed",
+                )
+                fc_item_done = ResponseOutputItemDoneEvent(
+                    sequence_number=next(seq),
+                    output_index=next_output_index,
+                    item=fc_done_item,
+                )
+                yield f"event: response.output_item.done\ndata: {fc_item_done.model_dump_json()}\n\n"
+
+                function_call_items.append(fc_done_item)
+                next_output_index += 1
+            continue
+
+        accumulated_text += chunk.text
+        last_stats = chunk.stats or last_stats
+
+        # response.output_text.delta
+        delta_event = ResponseTextDeltaEvent(
+            sequence_number=next(seq),
+            item_id=item_id,
+            output_index=0,
+            content_index=0,
+            delta=chunk.text,
+        )
+        yield f"event: response.output_text.delta\ndata: {delta_event.model_dump_json()}\n\n"
+
+    # response.output_text.done
+    text_done = ResponseTextDoneEvent(
+        sequence_number=next(seq),
+        item_id=item_id,
+        output_index=0,
+        content_index=0,
+        text=accumulated_text,
+    )
+    yield f"event: response.output_text.done\ndata: {text_done.model_dump_json()}\n\n"
+
+    # response.content_part.done
+    final_part = ResponseOutputText(text=accumulated_text)
+    part_done = ResponseContentPartDoneEvent(
+        sequence_number=next(seq),
+        item_id=item_id,
+        output_index=0,
+        content_index=0,
+        part=final_part,
+    )
+    yield f"event: response.content_part.done\ndata: {part_done.model_dump_json()}\n\n"
+
+    # response.output_item.done
+    final_message_item = ResponseMessageItem(
+        id=item_id,
+        content=[ResponseOutputText(text=accumulated_text)],
+        status="completed",
+    )
+    item_done = ResponseOutputItemDoneEvent(
+        sequence_number=next(seq), output_index=0, item=final_message_item
+    )
+    yield f"event: response.output_item.done\ndata: {item_done.model_dump_json()}\n\n"
+
+    # Create usage from stats if available
+    usage = None
+    if last_stats is not None:
+        usage = ResponseUsage(
+            input_tokens=last_stats.prompt_tokens,
+            output_tokens=last_stats.generation_tokens,
+            total_tokens=last_stats.prompt_tokens + last_stats.generation_tokens,
+        )
+
+    # response.completed
+    output: list[ResponseItem] = [final_message_item]
+    output.extend(function_call_items)
+    final_response = ResponsesResponse(
+        id=response_id,
+        model=model,
+        status="completed",
+        output=output,
+        output_text=accumulated_text,
+        usage=usage,
+    )
+    completed_event = ResponseCompletedEvent(
+        sequence_number=next(seq), response=final_response
+    )
+    yield f"event: response.completed\ndata: {completed_event.model_dump_json()}\n\n"

src/exo/master/api.py

@@ -1,9 +1,12 @@
 import base64
 import contextlib
 import json
+import random
 import time
-from collections.abc import AsyncGenerator
+from collections.abc import AsyncGenerator, Awaitable, Callable
+from datetime import datetime, timezone
 from http import HTTPStatus
+from pathlib import Path
 from typing import Annotated, Literal, cast
 from uuid import uuid4
 
@@ -19,28 +22,50 @@ from hypercorn.config import Config
 from hypercorn.typing import ASGIFramework
 from loguru import logger
 
+from exo.master.adapters.chat_completions import (
+    chat_request_to_text_generation,
+    collect_chat_response,
+    generate_chat_stream,
+)
+from exo.master.adapters.claude import (
+    claude_request_to_text_generation,
+    collect_claude_response,
+    generate_claude_stream,
+)
+from exo.master.adapters.responses import (
+    collect_responses_response,
+    generate_responses_stream,
+    responses_request_to_text_generation,
+)
 from exo.master.image_store import ImageStore
 from exo.master.placement import place_instance as get_instance_placements
 from exo.shared.apply import apply
 from exo.shared.constants import (
+    DASHBOARD_DIR,
     EXO_IMAGE_CACHE_DIR,
     EXO_MAX_CHUNK_SIZE,
+    EXO_TRACING_CACHE_DIR,
 )
 from exo.shared.election import ElectionMessage
 from exo.shared.logging import InterceptLogger
 from exo.shared.models.model_cards import (
-    MODEL_CARDS,
     ModelCard,
     ModelId,
+    delete_custom_card,
+    get_model_cards,
+    is_custom_card,
 )
+from exo.shared.tracing import TraceEvent, compute_stats, export_trace, load_trace_file
 from exo.shared.types.api import (
+    AddCustomModelParams,
     AdvancedImageParams,
+    BenchChatCompletionRequest,
     BenchChatCompletionResponse,
-    BenchChatCompletionTaskParams,
     BenchImageGenerationResponse,
     BenchImageGenerationTaskParams,
     ChatCompletionChoice,
     ChatCompletionMessage,
+    ChatCompletionRequest,
     ChatCompletionResponse,
     CreateInstanceParams,
     CreateInstanceResponse,
@@ -50,8 +75,9 @@ from exo.shared.types.api import (
     ErrorResponse,
     FinishReason,
     GenerationStats,
+    HuggingFaceSearchResult,
     ImageData,
-    ImageEditsInternalParams,
+    ImageEditsTaskParams,
     ImageGenerationResponse,
     ImageGenerationStats,
     ImageGenerationTaskParams,
@@ -64,10 +90,14 @@ from exo.shared.types.api import (
     PlacementPreviewResponse,
     StartDownloadParams,
     StartDownloadResponse,
-    StreamingChoiceResponse,
-    StreamOptions,
     ToolCall,
-    Usage,
+    TraceCategoryStats,
+    TraceEventResponse,
+    TraceListItem,
+    TraceListResponse,
+    TraceRankStats,
+    TraceResponse,
+    TraceStatsResponse,
 )
 from exo.shared.types.chunks import (
     ErrorChunk,
@@ -76,8 +106,11 @@ from exo.shared.types.chunks import (
     TokenChunk,
     ToolCallChunk,
 )
+from exo.shared.types.claude_api import (
+    ClaudeMessagesRequest,
+    ClaudeMessagesResponse,
+)
 from exo.shared.types.commands import (
-    ChatCompletion,
     Command,
     CreateInstance,
     DeleteDownload,
@@ -91,6 +124,7 @@ from exo.shared.types.commands import (
     SendInputChunk,
     StartDownload,
     TaskFinished,
+    TextGeneration,
 )
 from exo.shared.types.common import CommandId, Id, NodeId, SessionId
 from exo.shared.types.events import (
@@ -98,15 +132,18 @@ from exo.shared.types.events import (
     Event,
     ForwarderEvent,
     IndexedEvent,
+    TracesMerged,
 )
 from exo.shared.types.memory import Memory
+from exo.shared.types.openai_responses import (
+    ResponsesRequest,
+    ResponsesResponse,
+)
 from exo.shared.types.state import State
-from exo.shared.types.tasks import ChatCompletionTaskParams
 from exo.shared.types.worker.instances import Instance, InstanceId, InstanceMeta
 from exo.shared.types.worker.shards import Sharding
 from exo.utils.banner import print_startup_banner
 from exo.utils.channels import Receiver, Sender, channel
-from exo.utils.dashboard_path import find_dashboard
 from exo.utils.event_buffer import OrderedBuffer
 
 
@@ -114,36 +151,13 @@ def _format_to_content_type(image_format: Literal["png", "jpeg", "webp"] | None)
     return f"image/{image_format or 'png'}"
 
 
-def chunk_to_response(
-    chunk: TokenChunk | ToolCallChunk,
-    command_id: CommandId,
-    usage: Usage | None,
-) -> ChatCompletionResponse:
-    return ChatCompletionResponse(
-        id=command_id,
-        created=int(time.time()),
-        model=chunk.model,
-        choices=[
-            StreamingChoiceResponse(
-                index=0,
-                delta=ChatCompletionMessage(role="assistant", content=chunk.text)
-                if isinstance(chunk, TokenChunk)
-                else ChatCompletionMessage(
-                    role="assistant",
-                    tool_calls=[
-                        ToolCall(
-                            id=str(uuid4()),
-                            index=i,
-                            function=tool,
-                        )
-                        for i, tool in enumerate(chunk.tool_calls)
-                    ],
-                ),
-                finish_reason=chunk.finish_reason,
-            )
-        ],
-        usage=usage,
-    )
+def _ensure_seed(params: AdvancedImageParams | None) -> AdvancedImageParams:
+    """Ensure advanced params has a seed set for distributed consistency."""
+    if params is None:
+        return AdvancedImageParams(seed=random.randint(0, 2**32 - 1))
+    if params.seed is None:
+        return params.model_copy(update={"seed": random.randint(0, 2**32 - 1)})
+    return params
 
 
 class API:
@@ -176,6 +190,15 @@ class API:
         self.paused_ev: anyio.Event = anyio.Event()
 
         self.app = FastAPI()
+
+        @self.app.middleware("http")
+        async def _log_requests(  # pyright: ignore[reportUnusedFunction]
+            request: Request,
+            call_next: Callable[[Request], Awaitable[StreamingResponse]],
+        ) -> StreamingResponse:
+            logger.debug(f"API request: {request.method} {request.url.path}")
+            return await call_next(request)
+
         self._setup_exception_handlers()
         self._setup_cors()
         self._setup_routes()
@@ -183,13 +206,13 @@ class API:
         self.app.mount(
             "/",
             StaticFiles(
-                directory=find_dashboard(),
+                directory=DASHBOARD_DIR,
                 html=True,
             ),
             name="dashboard",
         )
 
-        self._chat_completion_queues: dict[
+        self._text_generation_queues: dict[
             CommandId, Sender[TokenChunk | ErrorChunk | ToolCallChunk]
         ] = {}
         self._image_generation_queues: dict[
@@ -203,7 +226,7 @@ class API:
         self.state = State()
         self.session_id = new_session_id
         self.event_buffer = OrderedBuffer[Event]()
-        self._chat_completion_queues = {}
+        self._text_generation_queues = {}
         self._image_generation_queues = {}
         self.unpause(result_clock)
 
@@ -248,6 +271,9 @@ class API:
         self.app.delete("/instance/{instance_id}")(self.delete_instance)
         self.app.get("/models")(self.get_models)
         self.app.get("/v1/models")(self.get_models)
+        self.app.post("/models/add")(self.add_custom_model)
+        self.app.delete("/models/custom/{model_id:path}")(self.delete_custom_model)
+        self.app.get("/models/search")(self.search_models)
         self.app.post("/v1/chat/completions", response_model=None)(
             self.chat_completions
         )
@@ -260,10 +286,16 @@ class API:
         self.app.post("/bench/images/edits")(self.bench_image_edits)
         self.app.get("/images")(self.list_images)
         self.app.get("/images/{image_id}")(self.get_image)
+        self.app.post("/v1/messages", response_model=None)(self.claude_messages)
+        self.app.post("/v1/responses", response_model=None)(self.openai_responses)
         self.app.get("/state")(lambda: self.state)
         self.app.get("/events")(lambda: self._event_log)
         self.app.post("/download/start")(self.start_download)
         self.app.delete("/download/{node_id}/{model_id:path}")(self.delete_download)
+        self.app.get("/v1/traces")(self.list_traces)
+        self.app.get("/v1/traces/{task_id}")(self.get_trace)
+        self.app.get("/v1/traces/{task_id}/stats")(self.get_trace_stats)
+        self.app.get("/v1/traces/{task_id}/raw")(self.get_trace_raw)
 
     async def place_instance(self, payload: PlaceInstanceParams):
         command = PlaceInstance(
@@ -354,10 +386,7 @@ class API:
         if len(list(self.state.topology.list_nodes())) == 0:
             return PlacementPreviewResponse(previews=[])
 
-        cards = [card for card in MODEL_CARDS.values() if card.model_id == model_id]
-        if not cards:
-            raise HTTPException(status_code=404, detail=f"Model {model_id} not found")
-
+        model_card = await ModelCard.load(model_id)
         instance_combinations: list[tuple[Sharding, InstanceMeta, int]] = []
         for sharding in (Sharding.Pipeline, Sharding.Tensor):
             for instance_meta in (InstanceMeta.MlxRing, InstanceMeta.MlxJaccl):
@@ -372,96 +401,93 @@ class API:
         # TODO: PDD
         # instance_combinations.append((Sharding.PrefillDecodeDisaggregation, InstanceMeta.MlxRing, 1))
 
-        for model_card in cards:
-            for sharding, instance_meta, min_nodes in instance_combinations:
-                try:
-                    placements = get_instance_placements(
-                        PlaceInstance(
-                            model_card=model_card,
-                            sharding=sharding,
-                            instance_meta=instance_meta,
-                            min_nodes=min_nodes,
-                        ),
-                        node_memory=self.state.node_memory,
-                        node_network=self.state.node_network,
-                        topology=self.state.topology,
-                        current_instances=self.state.instances,
-                        required_nodes=required_nodes,
-                    )
-                except ValueError as exc:
-                    if (model_card.model_id, sharding, instance_meta, 0) not in seen:
-                        previews.append(
-                            PlacementPreview(
-                                model_id=model_card.model_id,
-                                sharding=sharding,
-                                instance_meta=instance_meta,
-                                instance=None,
-                                error=str(exc),
-                            )
-                        )
-                    seen.add((model_card.model_id, sharding, instance_meta, 0))
-                    continue
-
-                current_ids = set(self.state.instances.keys())
-                new_instances = [
-                    instance
-                    for instance_id, instance in placements.items()
-                    if instance_id not in current_ids
-                ]
-
-                if len(new_instances) != 1:
-                    if (model_card.model_id, sharding, instance_meta, 0) not in seen:
-                        previews.append(
-                            PlacementPreview(
-                                model_id=model_card.model_id,
-                                sharding=sharding,
-                                instance_meta=instance_meta,
-                                instance=None,
-                                error="Expected exactly one new instance from placement",
-                            )
-                        )
-                    seen.add((model_card.model_id, sharding, instance_meta, 0))
-                    continue
-
-                instance = new_instances[0]
-                shard_assignments = instance.shard_assignments
-                placement_node_ids = list(shard_assignments.node_to_runner.keys())
-
-                memory_delta_by_node: dict[str, int] = {}
-                if placement_node_ids:
-                    total_bytes = model_card.storage_size.in_bytes
-                    per_node = total_bytes // len(placement_node_ids)
-                    remainder = total_bytes % len(placement_node_ids)
-                    for index, node_id in enumerate(
-                        sorted(placement_node_ids, key=str)
-                    ):
-                        extra = 1 if index < remainder else 0
-                        memory_delta_by_node[str(node_id)] = per_node + extra
-
-                if (
-                    model_card.model_id,
-                    sharding,
-                    instance_meta,
-                    len(placement_node_ids),
-                ) not in seen:
+        for sharding, instance_meta, min_nodes in instance_combinations:
+            try:
+                placements = get_instance_placements(
+                    PlaceInstance(
+                        model_card=model_card,
+                        sharding=sharding,
+                        instance_meta=instance_meta,
+                        min_nodes=min_nodes,
+                    ),
+                    node_memory=self.state.node_memory,
+                    node_network=self.state.node_network,
+                    topology=self.state.topology,
+                    current_instances=self.state.instances,
+                    required_nodes=required_nodes,
+                )
+            except ValueError as exc:
+                if (model_card.model_id, sharding, instance_meta, 0) not in seen:
                     previews.append(
                         PlacementPreview(
                             model_id=model_card.model_id,
                             sharding=sharding,
                             instance_meta=instance_meta,
-                            instance=instance,
-                            memory_delta_by_node=memory_delta_by_node or None,
-                            error=None,
+                            instance=None,
+                            error=str(exc),
                         )
                     )
-                seen.add(
-                    (
-                        model_card.model_id,
-                        sharding,
-                        instance_meta,
-                        len(placement_node_ids),
+                seen.add((model_card.model_id, sharding, instance_meta, 0))
+                continue
+
+            current_ids = set(self.state.instances.keys())
+            new_instances = [
+                instance
+                for instance_id, instance in placements.items()
+                if instance_id not in current_ids
+            ]
+
+            if len(new_instances) != 1:
+                if (model_card.model_id, sharding, instance_meta, 0) not in seen:
+                    previews.append(
+                        PlacementPreview(
+                            model_id=model_card.model_id,
+                            sharding=sharding,
+                            instance_meta=instance_meta,
+                            instance=None,
+                            error="Expected exactly one new instance from placement",
+                        )
+                    )
+                seen.add((model_card.model_id, sharding, instance_meta, 0))
+                continue
+
+            instance = new_instances[0]
+            shard_assignments = instance.shard_assignments
+            placement_node_ids = list(shard_assignments.node_to_runner.keys())
+
+            memory_delta_by_node: dict[str, int] = {}
+            if placement_node_ids:
+                total_bytes = model_card.storage_size.in_bytes
+                per_node = total_bytes // len(placement_node_ids)
+                remainder = total_bytes % len(placement_node_ids)
+                for index, node_id in enumerate(sorted(placement_node_ids, key=str)):
+                    extra = 1 if index < remainder else 0
+                    memory_delta_by_node[str(node_id)] = per_node + extra
+
+            if (
+                model_card.model_id,
+                sharding,
+                instance_meta,
+                len(placement_node_ids),
+            ) not in seen:
+                previews.append(
+                    PlacementPreview(
+                        model_id=model_card.model_id,
+                        sharding=sharding,
+                        instance_meta=instance_meta,
+                        instance=instance,
+                        memory_delta_by_node=memory_delta_by_node or None,
+                        error=None,
                     )
                 )
+            seen.add(
+                (
+                    model_card.model_id,
+                    sharding,
+                    instance_meta,
+                    len(placement_node_ids),
+                )
+            )
 
         return PlacementPreviewResponse(previews=previews)
 
@@ -484,13 +510,15 @@ class API:
             instance_id=instance_id,
         )
 
-    async def _chat_chunk_stream(
+    async def _token_chunk_stream(
         self, command_id: CommandId
     ) -> AsyncGenerator[ErrorChunk | ToolCallChunk | TokenChunk, None]:
-        """Yield `TokenChunk`s for a given command until completion."""
+        """Yield chunks for a given command until completion.
 
+        This is the internal low-level stream used by all API adapters.
+        """
         try:
-            self._chat_completion_queues[command_id], recv = channel[
+            self._text_generation_queues[command_id], recv = channel[
                 ErrorChunk | ToolCallChunk | TokenChunk
             ]()
 
@@ -511,103 +539,10 @@ class API:
         finally:
             command = TaskFinished(finished_command_id=command_id)
             await self._send(command)
-            if command_id in self._chat_completion_queues:
-                del self._chat_completion_queues[command_id]
+            if command_id in self._text_generation_queues:
+                del self._text_generation_queues[command_id]
 
-    async def _generate_chat_stream(
-        self, command_id: CommandId, stream_options: StreamOptions | None = None
-    ) -> AsyncGenerator[str, None]:
-        """Generate chat completion stream as JSON strings."""
-        include_usage = stream_options.include_usage if stream_options else False
-
-        async for chunk in self._chat_chunk_stream(command_id):
-            assert not isinstance(chunk, ImageChunk)
-            if chunk.finish_reason == "error":
-                error_response = ErrorResponse(
-                    error=ErrorInfo(
-                        message=chunk.error_message or "Internal server error",
-                        type="InternalServerError",
-                        code=500,
-                    )
-                )
-                yield f"data: {error_response.model_dump_json()}\n\n"
-                yield "data: [DONE]\n\n"
-                return
-
-            usage = chunk.usage if include_usage else None
-
-            chunk_response: ChatCompletionResponse = chunk_to_response(
-                chunk, command_id, usage=usage
-            )
-            logger.debug(f"chunk_response: {chunk_response}")
-
-            yield f"data: {chunk_response.model_dump_json()}\n\n"
-
-            if chunk.finish_reason is not None:
-                yield "data: [DONE]\n\n"
-
-    async def _collect_chat_completion(
-        self, command_id: CommandId
-    ) -> ChatCompletionResponse:
-        """Collect all token chunks for a chat completion and return a single response."""
-
-        text_parts: list[str] = []
-        tool_calls: list[ToolCall] = []
-        model: ModelId | None = None
-        finish_reason: FinishReason | None = None
-        usage: Usage | None = None
-
-        async for chunk in self._chat_chunk_stream(command_id):
-            if isinstance(chunk, ErrorChunk):
-                raise HTTPException(
-                    status_code=500,
-                    detail=chunk.error_message or "Internal server error",
-                )
-
-            if model is None:
-                model = chunk.model
-
-            if isinstance(chunk, TokenChunk):
-                text_parts.append(chunk.text)
-
-            if isinstance(chunk, ToolCallChunk):
-                tool_calls.extend(
-                    ToolCall(
-                        id=str(uuid4()),
-                        index=i,
-                        function=tool,
-                    )
-                    for i, tool in enumerate(chunk.tool_calls)
-                )
-
-            if chunk.usage is not None:
-                usage = chunk.usage
-
-            if chunk.finish_reason is not None:
-                finish_reason = chunk.finish_reason
-
-        combined_text = "".join(text_parts)
-        assert model is not None
-
-        return ChatCompletionResponse(
-            id=command_id,
-            created=int(time.time()),
-            model=model,
-            choices=[
-                ChatCompletionChoice(
-                    index=0,
-                    message=ChatCompletionMessage(
-                        role="assistant",
-                        content=combined_text,
-                        tool_calls=tool_calls,
-                    ),
-                    finish_reason=finish_reason,
-                )
-            ],
-            usage=usage,
-        )
-
-    async def _collect_chat_completion_with_stats(
+    async def _collect_text_generation_with_stats(
         self, command_id: CommandId
     ) -> BenchChatCompletionResponse:
         text_parts: list[str] = []
@@ -617,7 +552,7 @@ class API:
 
         stats: GenerationStats | None = None
 
-        async for chunk in self._chat_chunk_stream(command_id):
+        async for chunk in self._token_chunk_stream(command_id):
             if chunk.finish_reason == "error":
                 raise HTTPException(
                     status_code=500,
@@ -656,7 +591,9 @@ class API:
                 ChatCompletionChoice(
                     index=0,
                     message=ChatCompletionMessage(
-                        role="assistant", content=combined_text, tool_calls=tool_calls
+                        role="assistant",
+                        content=combined_text,
+                        tool_calls=tool_calls if tool_calls else None,
                     ),
                     finish_reason=finish_reason,
                 )
@@ -671,56 +608,65 @@ class API:
         )
 
     async def chat_completions(
-        self, payload: ChatCompletionTaskParams
+        self, payload: ChatCompletionRequest
     ) -> ChatCompletionResponse | StreamingResponse:
-        """Handle chat completions, supporting both streaming and non-streaming responses."""
-        model_card = await ModelCard.load(ModelId(payload.model))
-        payload.model = model_card.model_id
-
-        if not any(
-            instance.shard_assignments.model_id == payload.model
-            for instance in self.state.instances.values()
-        ):
-            await self._trigger_notify_user_to_download_model(payload.model)
-            raise HTTPException(
-                status_code=404, detail=f"No instance found for model {payload.model}"
-            )
-
-        command = ChatCompletion(
-            request_params=payload,
+        """OpenAI Chat Completions API - adapter."""
+        task_params = chat_request_to_text_generation(payload)
+        resolved_model = await self._resolve_and_validate_text_model(
+            ModelId(task_params.model)
         )
+        task_params = task_params.model_copy(update={"model": resolved_model})
+
+        command = TextGeneration(task_params=task_params)
         await self._send(command)
+
         if payload.stream:
             return StreamingResponse(
-                self._generate_chat_stream(command.command_id, payload.stream_options),
+                generate_chat_stream(
+                    command.command_id,
+                    self._token_chunk_stream(command.command_id),
+                ),
                 media_type="text/event-stream",
             )
 
-        return await self._collect_chat_completion(command.command_id)
+        return await collect_chat_response(
+            command.command_id,
+            self._token_chunk_stream(command.command_id),
+        )
 
     async def bench_chat_completions(
-        self, payload: BenchChatCompletionTaskParams
+        self, payload: BenchChatCompletionRequest
     ) -> BenchChatCompletionResponse:
-        model_card = await ModelCard.load(ModelId(payload.model))
-        payload.model = model_card.model_id
+        task_params = chat_request_to_text_generation(payload)
+        resolved_model = await self._resolve_and_validate_text_model(
+            ModelId(task_params.model)
+        )
+        task_params = task_params.model_copy(update={"model": resolved_model})
 
-        if not any(
-            instance.shard_assignments.model_id == payload.model
-            for instance in self.state.instances.values()
-        ):
-            await self._trigger_notify_user_to_download_model(payload.model)
-            raise HTTPException(
-                status_code=404, detail=f"No instance found for model {payload.model}"
-            )
+        task_params = task_params.model_copy(update={"stream": False, "bench": True})
 
-        payload.stream = False
-
-        command = ChatCompletion(request_params=payload)
+        command = TextGeneration(task_params=task_params)
         await self._send(command)
 
-        response = await self._collect_chat_completion_with_stats(command.command_id)
+        response = await self._collect_text_generation_with_stats(command.command_id)
         return response
 
+    async def _resolve_and_validate_text_model(self, model_id: ModelId) -> ModelId:
+        """Validate a text model exists and return the resolved model ID.
+
+        Raises HTTPException 404 if no instance is found for the model.
+        """
+        if not any(
+            instance.shard_assignments.model_id == model_id
+            for instance in self.state.instances.values()
+        ):
+            await self._trigger_notify_user_to_download_model(model_id)
+            raise HTTPException(
+                status_code=404,
+                detail=f"No instance found for model {model_id}",
+            )
+        return model_id
+
     async def _validate_image_model(self, model: ModelId) -> ModelId:
         """Validate model exists and return resolved model ID.
 
@@ -773,9 +719,12 @@ class API:
         with SSE-formatted events for partial and final images.
         """
         payload.model = await self._validate_image_model(ModelId(payload.model))
+        payload = payload.model_copy(
+            update={"advanced_params": _ensure_seed(payload.advanced_params)}
+        )
 
         command = ImageGeneration(
-            request_params=payload,
+            task_params=payload,
         )
         await self._send(command)
 
@@ -1021,9 +970,12 @@ class API:
 
         payload.stream = False
         payload.partial_images = 0
+        payload = payload.model_copy(
+            update={"advanced_params": _ensure_seed(payload.advanced_params)}
+        )
 
         command = ImageGeneration(
-            request_params=payload,
+            task_params=payload,
         )
         await self._send(command)
 
@@ -1052,6 +1004,7 @@ class API:
     ) -> ImageEdits:
         """Prepare and send an image edits command with chunked image upload."""
         resolved_model = await self._validate_image_model(model)
+        advanced_params = _ensure_seed(advanced_params)
 
         image_content = await image.read()
         image_data = base64.b64encode(image_content).decode("utf-8")
@@ -1065,7 +1018,7 @@ class API:
         total_chunks = len(data_chunks)
 
         command = ImageEdits(
-            request_params=ImageEditsInternalParams(
+            task_params=ImageEditsTaskParams(
                 image_data="",
                 total_input_chunks=total_chunks,
                 prompt=prompt,
@@ -1209,6 +1162,62 @@ class API:
             response_format=response_format,
         )
 
+    async def claude_messages(
+        self, payload: ClaudeMessagesRequest
+    ) -> ClaudeMessagesResponse | StreamingResponse:
+        """Claude Messages API - adapter."""
+        task_params = claude_request_to_text_generation(payload)
+        resolved_model = await self._resolve_and_validate_text_model(
+            ModelId(task_params.model)
+        )
+        task_params = task_params.model_copy(update={"model": resolved_model})
+
+        command = TextGeneration(task_params=task_params)
+        await self._send(command)
+
+        if payload.stream:
+            return StreamingResponse(
+                generate_claude_stream(
+                    command.command_id,
+                    payload.model,
+                    self._token_chunk_stream(command.command_id),
+                ),
+                media_type="text/event-stream",
+            )
+
+        return await collect_claude_response(
+            command.command_id,
+            payload.model,
+            self._token_chunk_stream(command.command_id),
+        )
+
+    async def openai_responses(
+        self, payload: ResponsesRequest
+    ) -> ResponsesResponse | StreamingResponse:
+        """OpenAI Responses API."""
+        task_params = responses_request_to_text_generation(payload)
+        resolved_model = await self._resolve_and_validate_text_model(task_params.model)
+        task_params = task_params.model_copy(update={"model": resolved_model})
+
+        command = TextGeneration(task_params=task_params)
+        await self._send(command)
+
+        if payload.stream:
+            return StreamingResponse(
+                generate_responses_stream(
+                    command.command_id,
+                    payload.model,
+                    self._token_chunk_stream(command.command_id),
+                ),
+                media_type="text/event-stream",
+            )
+
+        return await collect_responses_response(
+            command.command_id,
+            payload.model,
+            self._token_chunk_stream(command.command_id),
+        )
+
     def _calculate_total_available_memory(self) -> Memory:
         """Calculate total available memory across all nodes in bytes."""
         total_available = Memory()
@@ -1231,11 +1240,70 @@ class API:
                     storage_size_megabytes=int(card.storage_size.in_mb),
                     supports_tensor=card.supports_tensor,
                     tasks=[task.value for task in card.tasks],
+                    is_custom=is_custom_card(card.model_id),
+                    family=card.family,
+                    quantization=card.quantization,
+                    base_model=card.base_model,
+                    capabilities=card.capabilities,
                 )
-                for card in MODEL_CARDS.values()
+                for card in await get_model_cards()
             ]
         )
 
+    async def add_custom_model(self, payload: AddCustomModelParams) -> ModelListModel:
+        """Fetch a model from HuggingFace and save as a custom model card."""
+        try:
+            card = await ModelCard.fetch_from_hf(payload.model_id)
+        except Exception as exc:
+            raise HTTPException(
+                status_code=400, detail=f"Failed to fetch model: {exc}"
+            ) from exc
+
+        return ModelListModel(
+            id=card.model_id,
+            hugging_face_id=card.model_id,
+            name=card.model_id.short(),
+            description="",
+            tags=[],
+            storage_size_megabytes=int(card.storage_size.in_mb),
+            supports_tensor=card.supports_tensor,
+            tasks=[task.value for task in card.tasks],
+            is_custom=True,
+        )
+
+    async def delete_custom_model(self, model_id: ModelId) -> JSONResponse:
+        """Delete a user-added custom model card."""
+        deleted = await delete_custom_card(model_id)
+        if not deleted:
+            raise HTTPException(status_code=404, detail="Custom model card not found")
+        return JSONResponse(
+            {"message": "Model card deleted", "model_id": str(model_id)}
+        )
+
+    async def search_models(
+        self, query: str = "", limit: int = 20
+    ) -> list[HuggingFaceSearchResult]:
+        """Search HuggingFace Hub for mlx-community models."""
+        from huggingface_hub import list_models
+
+        results = list_models(
+            search=query or None,
+            author="mlx-community",
+            sort="downloads",
+            limit=limit,
+        )
+        return [
+            HuggingFaceSearchResult(
+                id=m.id,
+                author=m.author or "",
+                downloads=m.downloads or 0,
+                likes=m.likes or 0,
+                last_modified=str(m.last_modified or ""),
+                tags=list(m.tags or []),
+            )
+            for m in results
+        ]
+
     async def run(self):
         cfg = Config()
         cfg.bind = f"0.0.0.0:{self.port}"
@@ -1281,14 +1349,32 @@ class API:
                                 self._image_generation_queues.pop(
                                     event.command_id, None
                                 )
-                        if queue := self._chat_completion_queues.get(
+                        if queue := self._text_generation_queues.get(
                             event.command_id, None
                         ):
                             assert not isinstance(event.chunk, ImageChunk)
                             try:
                                 await queue.send(event.chunk)
                             except BrokenResourceError:
-                                self._chat_completion_queues.pop(event.command_id, None)
+                                self._text_generation_queues.pop(event.command_id, None)
+
+                    if isinstance(event, TracesMerged):
+                        self._save_merged_trace(event)
+
+    def _save_merged_trace(self, event: TracesMerged) -> None:
+        traces = [
+            TraceEvent(
+                name=t.name,
+                start_us=t.start_us,
+                duration_us=t.duration_us,
+                rank=t.rank,
+                category=t.category,
+            )
+            for t in event.traces
+        ]
+        output_path = EXO_TRACING_CACHE_DIR / f"trace_{event.task_id}.json"
+        export_trace(traces, output_path)
+        logger.debug(f"Saved merged trace to {output_path}")
 
     async def _pause_on_new_election(self):
         with self.election_receiver as ems:
@@ -1336,3 +1422,103 @@ class API:
         )
         await self._send_download(command)
         return DeleteDownloadResponse(command_id=command.command_id)
+
+    def _get_trace_path(self, task_id: str) -> Path:
+        return EXO_TRACING_CACHE_DIR / f"trace_{task_id}.json"
+
+    async def list_traces(self) -> TraceListResponse:
+        traces: list[TraceListItem] = []
+
+        for trace_file in sorted(
+            EXO_TRACING_CACHE_DIR.glob("trace_*.json"),
+            key=lambda p: p.stat().st_mtime,
+            reverse=True,
+        ):
+            # Extract task_id from filename (trace_{task_id}.json)
+            task_id = trace_file.stem.removeprefix("trace_")
+            stat = trace_file.stat()
+            created_at = datetime.fromtimestamp(
+                stat.st_mtime, tz=timezone.utc
+            ).isoformat()
+            traces.append(
+                TraceListItem(
+                    task_id=task_id,
+                    created_at=created_at,
+                    file_size=stat.st_size,
+                )
+            )
+
+        return TraceListResponse(traces=traces)
+
+    async def get_trace(self, task_id: str) -> TraceResponse:
+        trace_path = self._get_trace_path(task_id)
+
+        if not trace_path.exists():
+            raise HTTPException(status_code=404, detail=f"Trace not found: {task_id}")
+
+        trace_events = load_trace_file(trace_path)
+
+        return TraceResponse(
+            task_id=task_id,
+            traces=[
+                TraceEventResponse(
+                    name=event.name,
+                    start_us=event.start_us,
+                    duration_us=event.duration_us,
+                    rank=event.rank,
+                    category=event.category,
+                )
+                for event in trace_events
+            ],
+        )
+
+    async def get_trace_stats(self, task_id: str) -> TraceStatsResponse:
+        trace_path = self._get_trace_path(task_id)
+
+        if not trace_path.exists():
+            raise HTTPException(status_code=404, detail=f"Trace not found: {task_id}")
+
+        trace_events = load_trace_file(trace_path)
+        stats = compute_stats(trace_events)
+
+        return TraceStatsResponse(
+            task_id=task_id,
+            total_wall_time_us=stats.total_wall_time_us,
+            by_category={
+                category: TraceCategoryStats(
+                    total_us=cat_stats.total_us,
+                    count=cat_stats.count,
+                    min_us=cat_stats.min_us,
+                    max_us=cat_stats.max_us,
+                    avg_us=cat_stats.avg_us,
+                )
+                for category, cat_stats in stats.by_category.items()
+            },
+            by_rank={
+                rank: TraceRankStats(
+                    by_category={
+                        category: TraceCategoryStats(
+                            total_us=cat_stats.total_us,
+                            count=cat_stats.count,
+                            min_us=cat_stats.min_us,
+                            max_us=cat_stats.max_us,
+                            avg_us=cat_stats.avg_us,
+                        )
+                        for category, cat_stats in rank_stats.items()
+                    }
+                )
+                for rank, rank_stats in stats.by_rank.items()
+            },
+        )
+
+    async def get_trace_raw(self, task_id: str) -> FileResponse:
+        trace_path = self._get_trace_path(task_id)
+
+        if not trace_path.exists():
+            raise HTTPException(status_code=404, detail=f"Trace not found: {task_id}")
+
+        return FileResponse(
+            path=trace_path,
+            media_type="application/json",
+            filename=f"trace_{task_id}.json",
+        )

src/exo/master/main.py

@@ -11,8 +11,8 @@ from exo.master.placement import (
     place_instance,
 )
 from exo.shared.apply import apply
+from exo.shared.constants import EXO_TRACING_ENABLED
 from exo.shared.types.commands import (
-    ChatCompletion,
     CreateInstance,
     DeleteInstance,
     ForwarderCommand,
@@ -23,6 +23,7 @@ from exo.shared.types.commands import (
     SendInputChunk,
     TaskFinished,
     TestCommand,
+    TextGeneration,
 )
 from exo.shared.types.common import CommandId, NodeId, SessionId
 from exo.shared.types.events import (
@@ -35,11 +36,11 @@ from exo.shared.types.events import (
     NodeTimedOut,
     TaskCreated,
     TaskDeleted,
+    TraceEventData,
+    TracesCollected,
+    TracesMerged,
 )
 from exo.shared.types.state import State
-from exo.shared.types.tasks import (
-    ChatCompletion as ChatCompletionTask,
-)
 from exo.shared.types.tasks import (
     ImageEdits as ImageEditsTask,
 )
@@ -50,6 +51,9 @@ from exo.shared.types.tasks import (
     TaskId,
     TaskStatus,
 )
+from exo.shared.types.tasks import (
+    TextGeneration as TextGenerationTask,
+)
 from exo.shared.types.worker.instances import InstanceId
 from exo.utils.channels import Receiver, Sender, channel
 from exo.utils.event_buffer import MultiSourceBuffer
@@ -86,6 +90,8 @@ class Master:
         self._multi_buffer = MultiSourceBuffer[NodeId, Event]()
         # TODO: not have this
         self._event_log: list[Event] = []
+        self._pending_traces: dict[TaskId, dict[int, list[TraceEventData]]] = {}
+        self._expected_ranks: dict[TaskId, set[int]] = {}
 
     async def run(self):
         logger.info("Starting Master")
@@ -117,11 +123,11 @@ class Master:
                     match command:
                         case TestCommand():
                             pass
-                        case ChatCompletion():
+                        case TextGeneration():
                             for instance in self.state.instances.values():
                                 if (
                                     instance.shard_assignments.model_id
-                                    == command.request_params.model
+                                    == command.task_params.model
                                 ):
                                     task_count = sum(
                                         1
@@ -134,7 +140,7 @@ class Master:
 
                             if not instance_task_counts:
                                 raise ValueError(
-                                    f"No instance found for model {command.request_params.model}"
+                                    f"No instance found for model {command.task_params.model}"
                                 )
 
                             available_instance_ids = sorted(
@@ -148,12 +154,12 @@ class Master:
                             generated_events.append(
                                 TaskCreated(
                                     task_id=task_id,
-                                    task=ChatCompletionTask(
+                                    task=TextGenerationTask(
                                         task_id=task_id,
                                         command_id=command.command_id,
                                         instance_id=available_instance_ids[0],
                                         task_status=TaskStatus.Pending,
-                                        task_params=command.request_params,
+                                        task_params=command.task_params,
                                     ),
                                 )
                             )
@@ -163,7 +169,7 @@ class Master:
                             for instance in self.state.instances.values():
                                 if (
                                     instance.shard_assignments.model_id
-                                    == command.request_params.model
+                                    == command.task_params.model
                                 ):
                                     task_count = sum(
                                         1
@@ -176,7 +182,7 @@ class Master:
 
                             if not instance_task_counts:
                                 raise ValueError(
-                                    f"No instance found for model {command.request_params.model}"
+                                    f"No instance found for model {command.task_params.model}"
                                 )
 
                             available_instance_ids = sorted(
@@ -187,25 +193,37 @@ class Master:
                             )
 
                             task_id = TaskId()
+                            selected_instance_id = available_instance_ids[0]
                             generated_events.append(
                                 TaskCreated(
                                     task_id=task_id,
                                     task=ImageGenerationTask(
                                         task_id=task_id,
                                         command_id=command.command_id,
-                                        instance_id=available_instance_ids[0],
+                                        instance_id=selected_instance_id,
                                         task_status=TaskStatus.Pending,
-                                        task_params=command.request_params,
+                                        task_params=command.task_params,
                                     ),
                                 )
                             )
 
                             self.command_task_mapping[command.command_id] = task_id
+
+                            if EXO_TRACING_ENABLED:
+                                selected_instance = self.state.instances.get(
+                                    selected_instance_id
+                                )
+                                if selected_instance:
+                                    ranks = set(
+                                        shard.device_rank
+                                        for shard in selected_instance.shard_assignments.runner_to_shard.values()
+                                    )
+                                    self._expected_ranks[task_id] = ranks
                         case ImageEdits():
                             for instance in self.state.instances.values():
                                 if (
                                     instance.shard_assignments.model_id
-                                    == command.request_params.model
+                                    == command.task_params.model
                                 ):
                                     task_count = sum(
                                         1
@@ -218,7 +236,7 @@ class Master:
 
                             if not instance_task_counts:
                                 raise ValueError(
-                                    f"No instance found for model {command.request_params.model}"
+                                    f"No instance found for model {command.task_params.model}"
                                 )
 
                             available_instance_ids = sorted(
@@ -229,20 +247,32 @@ class Master:
                             )
 
                             task_id = TaskId()
+                            selected_instance_id = available_instance_ids[0]
                             generated_events.append(
                                 TaskCreated(
                                     task_id=task_id,
                                     task=ImageEditsTask(
                                         task_id=task_id,
                                         command_id=command.command_id,
-                                        instance_id=available_instance_ids[0],
+                                        instance_id=selected_instance_id,
                                         task_status=TaskStatus.Pending,
-                                        task_params=command.request_params,
+                                        task_params=command.task_params,
                                     ),
                                 )
                             )
 
                             self.command_task_mapping[command.command_id] = task_id
+
+                            if EXO_TRACING_ENABLED:
+                                selected_instance = self.state.instances.get(
+                                    selected_instance_id
+                                )
+                                if selected_instance:
+                                    ranks = set(
+                                        shard.device_rank
+                                        for shard in selected_instance.shard_assignments.runner_to_shard.values()
+                                    )
+                                    self._expected_ranks[task_id] = ranks
                         case DeleteInstance():
                             placement = delete_instance(command, self.state.instances)
                             transition_events = get_transition_events(
@@ -335,6 +365,10 @@ class Master:
                     local_event.origin,
                 )
                 for event in self._multi_buffer.drain():
+                    if isinstance(event, TracesCollected):
+                        await self._handle_traces_collected(event)
+                        continue
+
                     logger.debug(f"Master indexing event: {str(event)[:100]}")
                     indexed = IndexedEvent(event=event, idx=len(self._event_log))
                     self.state = apply(self.state, indexed)
@@ -373,3 +407,29 @@ class Master:
                 event=event.event,
             )
         )
+
+    async def _handle_traces_collected(self, event: TracesCollected) -> None:
+        task_id = event.task_id
+        if task_id not in self._pending_traces:
+            self._pending_traces[task_id] = {}
+        self._pending_traces[task_id][event.rank] = event.traces
+
+        if (
+            task_id in self._expected_ranks
+            and set(self._pending_traces[task_id].keys())
+            >= self._expected_ranks[task_id]
+        ):
+            await self._merge_and_save_traces(task_id)
+
+    async def _merge_and_save_traces(self, task_id: TaskId) -> None:
+        all_trace_data: list[TraceEventData] = []
+        for trace_data in self._pending_traces[task_id].values():
+            all_trace_data.extend(trace_data)
+
+        await self.event_sender.send(
+            TracesMerged(task_id=task_id, traces=all_trace_data)
+        )
+
+        del self._pending_traces[task_id]
+        if task_id in self._expected_ranks:
+            del self._expected_ranks[task_id]

src/exo/master/placement_utils.py

@@ -10,6 +10,7 @@ from exo.shared.types.profiling import MemoryUsage, NodeNetworkInfo
 from exo.shared.types.topology import Cycle, RDMAConnection, SocketConnection
 from exo.shared.types.worker.runners import RunnerId, ShardAssignments
 from exo.shared.types.worker.shards import (
+    CfgShardMetadata,
     PipelineShardMetadata,
     Sharding,
     ShardMetadata,
@@ -74,40 +75,43 @@ def allocate_layers_proportionally(
     return result
 
 
-def get_shard_assignments_for_pipeline_parallel(
-    model_card: ModelCard,
-    cycle: Cycle,
-    node_memory: Mapping[NodeId, MemoryUsage],
-):
+def _validate_cycle(cycle: Cycle) -> None:
     if not cycle.node_ids:
         raise ValueError("Cannot create shard assignments for empty node cycle")
 
-    cycle_memory = sum(
-        (node_memory[node_id].ram_available for node_id in cycle.node_ids),
+
+def _compute_total_memory(
+    node_ids: list[NodeId],
+    node_memory: Mapping[NodeId, MemoryUsage],
+) -> Memory:
+    total_memory = sum(
+        (node_memory[node_id].ram_available for node_id in node_ids),
         start=Memory(),
     )
-    if cycle_memory.in_bytes == 0:
+    if total_memory.in_bytes == 0:
         raise ValueError("Cannot create shard assignments: total available memory is 0")
+    return total_memory
 
-    total_layers = model_card.n_layers
-    world_size = len(cycle)
-    runner_to_shard: dict[RunnerId, ShardMetadata] = {}
-    node_to_runner: dict[NodeId, RunnerId] = {}
 
+def _allocate_and_validate_layers(
+    node_ids: list[NodeId],
+    node_memory: Mapping[NodeId, MemoryUsage],
+    total_memory: Memory,
+    model_card: ModelCard,
+) -> list[int]:
     layer_allocations = allocate_layers_proportionally(
-        total_layers=total_layers,
+        total_layers=model_card.n_layers,
         memory_fractions=[
-            node_memory[node_id].ram_available.in_bytes / cycle_memory.in_bytes
-            for node_id in cycle.node_ids
+            node_memory[node_id].ram_available.in_bytes / total_memory.in_bytes
+            for node_id in node_ids
         ],
     )
 
-    # Validate each node has sufficient memory for its assigned layers
-    memory_per_layer = model_card.storage_size.in_bytes / total_layers
-    for i, (node_id, node_layers) in enumerate(
-        zip(cycle.node_ids, layer_allocations, strict=True)
-    ):
-        required_memory = node_layers * memory_per_layer
+    total_storage_bytes = model_card.storage_size.in_bytes
+    total_layers = model_card.n_layers
+    for i, node_id in enumerate(node_ids):
+        node_layers = layer_allocations[i]
+        required_memory = (total_storage_bytes * node_layers) // total_layers
         available_memory = node_memory[node_id].ram_available.in_bytes
         if required_memory > available_memory:
             raise ValueError(
@@ -116,32 +120,125 @@ def get_shard_assignments_for_pipeline_parallel(
                 f"but only has {available_memory / (1024**3):.2f} GB available"
             )
 
-    layers_assigned = 0
-    for i, (node_id, node_layers) in enumerate(
-        zip(cycle.node_ids, layer_allocations, strict=True)
-    ):
-        runner_id = RunnerId()
+    return layer_allocations
 
-        shard = PipelineShardMetadata(
+
+def get_shard_assignments_for_pipeline_parallel(
+    model_card: ModelCard,
+    cycle: Cycle,
+    node_memory: Mapping[NodeId, MemoryUsage],
+) -> ShardAssignments:
+    """Create shard assignments for pipeline parallel execution."""
+    world_size = len(cycle)
+    use_cfg_parallel = model_card.uses_cfg and world_size >= 2 and world_size % 2 == 0
+
+    if use_cfg_parallel:
+        return _get_shard_assignments_for_cfg_parallel(model_card, cycle, node_memory)
+    else:
+        return _get_shard_assignments_for_pure_pipeline(model_card, cycle, node_memory)
+
+
+def _get_shard_assignments_for_cfg_parallel(
+    model_card: ModelCard,
+    cycle: Cycle,
+    node_memory: Mapping[NodeId, MemoryUsage],
+) -> ShardAssignments:
+    """Create shard assignments for CFG parallel execution.
+
+    CFG parallel runs two independent pipelines. Group 0 processes the positive
+    prompt, group 1 processes the negative prompt. The ring topology places
+    group 1's ranks in reverse order so both "last stages" are neighbors for
+    efficient CFG exchange.
+    """
+    _validate_cycle(cycle)
+
+    world_size = len(cycle)
+    cfg_world_size = 2
+    pipeline_world_size = world_size // cfg_world_size
+
+    # Allocate layers for one pipeline group (both groups run the same layers)
+    pipeline_node_ids = cycle.node_ids[:pipeline_world_size]
+    pipeline_memory = _compute_total_memory(pipeline_node_ids, node_memory)
+    layer_allocations = _allocate_and_validate_layers(
+        pipeline_node_ids, node_memory, pipeline_memory, model_card
+    )
+
+    # Ring topology: group 0 ascending [0,1,2,...], group 1 descending [...,2,1,0]
+    # This places both last stages as neighbors for CFG exchange.
+    position_to_cfg_pipeline = [(0, r) for r in range(pipeline_world_size)] + [
+        (1, r) for r in reversed(range(pipeline_world_size))
+    ]
+
+    runner_to_shard: dict[RunnerId, ShardMetadata] = {}
+    node_to_runner: dict[NodeId, RunnerId] = {}
+
+    for device_rank, node_id in enumerate(cycle.node_ids):
+        cfg_rank, pipeline_rank = position_to_cfg_pipeline[device_rank]
+        layers_before = sum(layer_allocations[:pipeline_rank])
+        node_layers = layer_allocations[pipeline_rank]
+
+        shard = CfgShardMetadata(
             model_card=model_card,
-            device_rank=i,
+            device_rank=device_rank,
             world_size=world_size,
-            start_layer=layers_assigned,
-            end_layer=layers_assigned + node_layers,
-            n_layers=total_layers,
+            start_layer=layers_before,
+            end_layer=layers_before + node_layers,
+            n_layers=model_card.n_layers,
+            cfg_rank=cfg_rank,
+            cfg_world_size=cfg_world_size,
+            pipeline_rank=pipeline_rank,
+            pipeline_world_size=pipeline_world_size,
         )
 
+        runner_id = RunnerId()
         runner_to_shard[runner_id] = shard
         node_to_runner[node_id] = runner_id
-        layers_assigned += node_layers
 
-    shard_assignments = ShardAssignments(
+    return ShardAssignments(
         model_id=model_card.model_id,
         runner_to_shard=runner_to_shard,
         node_to_runner=node_to_runner,
     )
 
-    return shard_assignments
+
+def _get_shard_assignments_for_pure_pipeline(
+    model_card: ModelCard,
+    cycle: Cycle,
+    node_memory: Mapping[NodeId, MemoryUsage],
+) -> ShardAssignments:
+    """Create shard assignments for pure pipeline execution."""
+    _validate_cycle(cycle)
+    total_memory = _compute_total_memory(cycle.node_ids, node_memory)
+
+    layer_allocations = _allocate_and_validate_layers(
+        cycle.node_ids, node_memory, total_memory, model_card
+    )
+
+    runner_to_shard: dict[RunnerId, ShardMetadata] = {}
+    node_to_runner: dict[NodeId, RunnerId] = {}
+
+    for pipeline_rank, node_id in enumerate(cycle.node_ids):
+        layers_before = sum(layer_allocations[:pipeline_rank])
+        node_layers = layer_allocations[pipeline_rank]
+
+        shard = PipelineShardMetadata(
+            model_card=model_card,
+            device_rank=pipeline_rank,
+            world_size=len(cycle),
+            start_layer=layers_before,
+            end_layer=layers_before + node_layers,
+            n_layers=model_card.n_layers,
+        )
+
+        runner_id = RunnerId()
+        runner_to_shard[runner_id] = shard
+        node_to_runner[node_id] = runner_id
+
+    return ShardAssignments(
+        model_id=model_card.model_id,
+        runner_to_shard=runner_to_shard,
+        node_to_runner=node_to_runner,
+    )
 
 
 def get_shard_assignments_for_tensor_parallel(

src/exo/master/tests/test_claude_api.py

@@ -0,0 +1,182 @@
+"""Tests for Claude Messages API conversion functions and types."""
+
+import pydantic
+import pytest
+
+from exo.master.adapters.claude import (
+    claude_request_to_text_generation,
+    finish_reason_to_claude_stop_reason,
+)
+from exo.shared.types.claude_api import (
+    ClaudeMessage,
+    ClaudeMessagesRequest,
+    ClaudeTextBlock,
+)
+from exo.shared.types.common import ModelId
+
+
+class TestFinishReasonToClaudeStopReason:
+    """Tests for finish_reason to Claude stop_reason mapping."""
+
+    def test_stop_maps_to_end_turn(self):
+        assert finish_reason_to_claude_stop_reason("stop") == "end_turn"
+
+    def test_length_maps_to_max_tokens(self):
+        assert finish_reason_to_claude_stop_reason("length") == "max_tokens"
+
+    def test_tool_calls_maps_to_tool_use(self):
+        assert finish_reason_to_claude_stop_reason("tool_calls") == "tool_use"
+
+    def test_function_call_maps_to_tool_use(self):
+        assert finish_reason_to_claude_stop_reason("function_call") == "tool_use"
+
+    def test_content_filter_maps_to_end_turn(self):
+        assert finish_reason_to_claude_stop_reason("content_filter") == "end_turn"
+
+    def test_none_returns_none(self):
+        assert finish_reason_to_claude_stop_reason(None) is None
+
+
+class TestClaudeRequestToInternal:
+    """Tests for converting Claude Messages API requests to TextGenerationTaskParams."""
+
+    def test_basic_request_conversion(self):
+        request = ClaudeMessagesRequest(
+            model=ModelId("claude-3-opus"),
+            max_tokens=100,
+            messages=[
+                ClaudeMessage(role="user", content="Hello"),
+            ],
+        )
+        params = claude_request_to_text_generation(request)
+
+        assert params.model == "claude-3-opus"
+        assert params.max_output_tokens == 100
+        assert isinstance(params.input, list)
+        assert len(params.input) == 1
+        assert params.input[0].role == "user"
+        assert params.input[0].content == "Hello"
+        assert params.instructions is None
+
+    def test_request_with_system_string(self):
+        request = ClaudeMessagesRequest(
+            model=ModelId("claude-3-opus"),
+            max_tokens=100,
+            system="You are a helpful assistant.",
+            messages=[
+                ClaudeMessage(role="user", content="Hello"),
+            ],
+        )
+        params = claude_request_to_text_generation(request)
+
+        assert params.instructions == "You are a helpful assistant."
+        assert isinstance(params.input, list)
+        assert len(params.input) == 1
+        assert params.input[0].role == "user"
+        assert params.input[0].content == "Hello"
+
+    def test_request_with_system_text_blocks(self):
+        request = ClaudeMessagesRequest(
+            model=ModelId("claude-3-opus"),
+            max_tokens=100,
+            system=[
+                ClaudeTextBlock(text="You are helpful. "),
+                ClaudeTextBlock(text="Be concise."),
+            ],
+            messages=[
+                ClaudeMessage(role="user", content="Hello"),
+            ],
+        )
+        params = claude_request_to_text_generation(request)
+
+        assert params.instructions == "You are helpful. Be concise."
+        assert isinstance(params.input, list)
+        assert len(params.input) == 1
+
+    def test_request_with_content_blocks(self):
+        request = ClaudeMessagesRequest(
+            model=ModelId("claude-3-opus"),
+            max_tokens=100,
+            messages=[
+                ClaudeMessage(
+                    role="user",
+                    content=[
+                        ClaudeTextBlock(text="First part. "),
+                        ClaudeTextBlock(text="Second part."),
+                    ],
+                ),
+            ],
+        )
+        params = claude_request_to_text_generation(request)
+
+        assert isinstance(params.input, list)
+        assert len(params.input) == 1
+        assert params.input[0].content == "First part. Second part."
+
+    def test_request_with_multi_turn_conversation(self):
+        request = ClaudeMessagesRequest(
+            model=ModelId("claude-3-opus"),
+            max_tokens=100,
+            messages=[
+                ClaudeMessage(role="user", content="Hello"),
+                ClaudeMessage(role="assistant", content="Hi there!"),
+                ClaudeMessage(role="user", content="How are you?"),
+            ],
+        )
+        params = claude_request_to_text_generation(request)
+
+        assert isinstance(params.input, list)
+        assert len(params.input) == 3
+        assert params.input[0].role == "user"
+        assert params.input[1].role == "assistant"
+        assert params.input[2].role == "user"
+
+    def test_request_with_optional_parameters(self):
+        request = ClaudeMessagesRequest(
+            model=ModelId("claude-3-opus"),
+            max_tokens=100,
+            messages=[ClaudeMessage(role="user", content="Hello")],
+            temperature=0.7,
+            top_p=0.9,
+            top_k=40,
+            stop_sequences=["STOP", "END"],
+            stream=True,
+        )
+        params = claude_request_to_text_generation(request)
+
+        assert params.temperature == 0.7
+        assert params.top_p == 0.9
+        assert params.top_k == 40
+        assert params.stop == ["STOP", "END"]
+        assert params.stream is True
+
+
+class TestClaudeMessagesRequestValidation:
+    """Tests for Claude Messages API request validation."""
+
+    def test_request_requires_model(self):
+        with pytest.raises(pydantic.ValidationError):
+            ClaudeMessagesRequest.model_validate(
+                {
+                    "max_tokens": 100,
+                    "messages": [{"role": "user", "content": "Hello"}],
+                }
+            )
+
+    def test_request_requires_max_tokens(self):
+        with pytest.raises(pydantic.ValidationError):
+            ClaudeMessagesRequest.model_validate(
+                {
+                    "model": "claude-3-opus",
+                    "messages": [{"role": "user", "content": "Hello"}],
+                }
+            )
+
+    def test_request_requires_messages(self):
+        with pytest.raises(pydantic.ValidationError):
+            ClaudeMessagesRequest.model_validate(
+                {
+                    "model": "claude-3-opus",
+                    "max_tokens": 100,
+                }
+            )

src/exo/master/tests/test_claude_tool_use.py

@@ -0,0 +1,265 @@
+"""Tests for Claude Messages API tool_use support in the adapter."""
+
+import json
+from collections.abc import AsyncGenerator
+from typing import Any, cast
+
+from exo.master.adapters.claude import collect_claude_response, generate_claude_stream
+from exo.shared.types.api import ToolCallItem
+from exo.shared.types.chunks import ErrorChunk, TokenChunk, ToolCallChunk
+from exo.shared.types.common import CommandId, ModelId
+
+
+async def _chunks_to_stream(
+    chunks: list[ErrorChunk | ToolCallChunk | TokenChunk],
+) -> AsyncGenerator[ErrorChunk | ToolCallChunk | TokenChunk, None]:
+    for chunk in chunks:
+        yield chunk
+
+
+MODEL = ModelId("test-model")
+COMMAND_ID = CommandId("cmd_test123")
+
+
+def _parse_sse_events(events: list[str]) -> list[dict[str, Any]]:
+    """Parse SSE event strings into JSON dicts."""
+    parsed: list[dict[str, Any]] = []
+    for event_str in events:
+        for line in event_str.strip().split("\n"):
+            if line.startswith("data: "):
+                parsed.append(cast(dict[str, Any], json.loads(line[6:])))
+    return parsed
+
+
+class TestCollectClaudeResponseToolUse:
+    """Tests for non-streaming tool_use response collection."""
+
+    async def test_tool_call_chunk_produces_tool_use_blocks(self):
+        chunks: list[ErrorChunk | ToolCallChunk | TokenChunk] = [
+            ToolCallChunk(
+                model=MODEL,
+                usage=None,
+                tool_calls=[
+                    ToolCallItem(
+                        name="get_weather",
+                        arguments='{"location": "San Francisco"}',
+                    )
+                ],
+            ),
+        ]
+        response = await collect_claude_response(
+            COMMAND_ID, "test-model", _chunks_to_stream(chunks)
+        )
+
+        assert response.stop_reason == "tool_use"
+        tool_blocks = [b for b in response.content if b.type == "tool_use"]
+        assert len(tool_blocks) == 1
+        block = tool_blocks[0]
+        assert block.type == "tool_use"
+        assert block.name == "get_weather"
+        assert block.input == {"location": "San Francisco"}
+        assert block.id.startswith("toolu_")
+
+    async def test_multiple_tool_calls(self):
+        chunks: list[ErrorChunk | ToolCallChunk | TokenChunk] = [
+            ToolCallChunk(
+                model=MODEL,
+                usage=None,
+                tool_calls=[
+                    ToolCallItem(
+                        name="get_weather",
+                        arguments='{"location": "SF"}',
+                    ),
+                    ToolCallItem(
+                        name="get_time",
+                        arguments='{"timezone": "PST"}',
+                    ),
+                ],
+            ),
+        ]
+        response = await collect_claude_response(
+            COMMAND_ID, "test-model", _chunks_to_stream(chunks)
+        )
+
+        assert response.stop_reason == "tool_use"
+        tool_blocks = [b for b in response.content if b.type == "tool_use"]
+        assert len(tool_blocks) == 2
+        assert tool_blocks[0].name == "get_weather"
+        assert tool_blocks[1].name == "get_time"
+
+    async def test_mixed_text_and_tool_use(self):
+        chunks: list[ErrorChunk | ToolCallChunk | TokenChunk] = [
+            TokenChunk(model=MODEL, text="Let me check ", token_id=1, usage=None),
+            TokenChunk(model=MODEL, text="the weather.", token_id=2, usage=None),
+            ToolCallChunk(
+                model=MODEL,
+                usage=None,
+                tool_calls=[
+                    ToolCallItem(
+                        name="get_weather",
+                        arguments='{"location": "NYC"}',
+                    )
+                ],
+            ),
+        ]
+        response = await collect_claude_response(
+            COMMAND_ID, "test-model", _chunks_to_stream(chunks)
+        )
+
+        assert response.stop_reason == "tool_use"
+        text_blocks = [b for b in response.content if b.type == "text"]
+        tool_blocks = [b for b in response.content if b.type == "tool_use"]
+        assert len(text_blocks) == 1
+        assert text_blocks[0].text == "Let me check the weather."
+        assert len(tool_blocks) == 1
+        assert tool_blocks[0].name == "get_weather"
+
+    async def test_no_content_produces_empty_text_block(self):
+        chunks: list[ErrorChunk | ToolCallChunk | TokenChunk] = []
+        response = await collect_claude_response(
+            COMMAND_ID, "test-model", _chunks_to_stream(chunks)
+        )
+        assert len(response.content) == 1
+        assert response.content[0].type == "text"
+
+
+class TestGenerateClaudeStreamToolUse:
+    """Tests for streaming tool_use event generation."""
+
+    async def test_tool_call_emits_tool_use_events(self):
+        chunks: list[ErrorChunk | ToolCallChunk | TokenChunk] = [
+            ToolCallChunk(
+                model=MODEL,
+                usage=None,
+                tool_calls=[
+                    ToolCallItem(
+                        name="get_weather",
+                        arguments='{"location": "SF"}',
+                    )
+                ],
+            ),
+        ]
+        events: list[str] = []
+        async for event in generate_claude_stream(
+            COMMAND_ID, "test-model", _chunks_to_stream(chunks)
+        ):
+            events.append(event)
+
+        parsed = _parse_sse_events(events)
+
+        # Find tool_use content_block_start
+        tool_starts = [
+            e
+            for e in parsed
+            if e.get("type") == "content_block_start"
+            and cast(dict[str, Any], e.get("content_block", {})).get("type")
+            == "tool_use"
+        ]
+        assert len(tool_starts) == 1
+        content_block = cast(dict[str, Any], tool_starts[0]["content_block"])
+        assert content_block["name"] == "get_weather"
+        assert content_block["input"] == {}
+        assert cast(str, content_block["id"]).startswith("toolu_")
+
+        # Find input_json_delta
+        json_deltas = [
+            e
+            for e in parsed
+            if e.get("type") == "content_block_delta"
+            and cast(dict[str, Any], e.get("delta", {})).get("type")
+            == "input_json_delta"
+        ]
+        assert len(json_deltas) == 1
+        delta = cast(dict[str, Any], json_deltas[0]["delta"])
+        assert json.loads(cast(str, delta["partial_json"])) == {"location": "SF"}
+
+        # Find message_delta with tool_use stop reason
+        msg_deltas = [e for e in parsed if e.get("type") == "message_delta"]
+        assert len(msg_deltas) == 1
+        assert cast(dict[str, Any], msg_deltas[0]["delta"])["stop_reason"] == "tool_use"
+
+    async def test_streaming_mixed_text_and_tool_use(self):
+        chunks: list[ErrorChunk | ToolCallChunk | TokenChunk] = [
+            TokenChunk(model=MODEL, text="Hello ", token_id=1, usage=None),
+            ToolCallChunk(
+                model=MODEL,
+                usage=None,
+                tool_calls=[
+                    ToolCallItem(
+                        name="search",
+                        arguments='{"query": "test"}',
+                    )
+                ],
+            ),
+        ]
+        events: list[str] = []
+        async for event in generate_claude_stream(
+            COMMAND_ID, "test-model", _chunks_to_stream(chunks)
+        ):
+            events.append(event)
+
+        parsed = _parse_sse_events(events)
+
+        # Should have text delta at index 0
+        text_deltas = [
+            e
+            for e in parsed
+            if e.get("type") == "content_block_delta"
+            and cast(dict[str, Any], e.get("delta", {})).get("type") == "text_delta"
+        ]
+        assert len(text_deltas) == 1
+        assert text_deltas[0]["index"] == 0
+        assert cast(dict[str, Any], text_deltas[0]["delta"])["text"] == "Hello "
+
+        # Tool block at index 1
+        tool_starts = [
+            e
+            for e in parsed
+            if e.get("type") == "content_block_start"
+            and cast(dict[str, Any], e.get("content_block", {})).get("type")
+            == "tool_use"
+        ]
+        assert len(tool_starts) == 1
+        assert tool_starts[0]["index"] == 1
+
+        # Stop reason should be tool_use
+        msg_deltas = [e for e in parsed if e.get("type") == "message_delta"]
+        assert cast(dict[str, Any], msg_deltas[0]["delta"])["stop_reason"] == "tool_use"
+
+    async def test_streaming_tool_block_stop_events(self):
+        chunks: list[ErrorChunk | ToolCallChunk | TokenChunk] = [
+            ToolCallChunk(
+                model=MODEL,
+                usage=None,
+                tool_calls=[
+                    ToolCallItem(name="fn1", arguments="{}"),
+                    ToolCallItem(name="fn2", arguments='{"a": 1}'),
+                ],
+            ),
+        ]
+        events: list[str] = []
+        async for event in generate_claude_stream(
+            COMMAND_ID, "test-model", _chunks_to_stream(chunks)
+        ):
+            events.append(event)
+
+        parsed = _parse_sse_events(events)
+
+        # Two tool block starts (at indices 1 and 2)
+        tool_starts = [
+            e
+            for e in parsed
+            if e.get("type") == "content_block_start"
+            and cast(dict[str, Any], e.get("content_block", {})).get("type")
+            == "tool_use"
+        ]
+        assert len(tool_starts) == 2
+        assert tool_starts[0]["index"] == 1
+        assert tool_starts[1]["index"] == 2
+
+        # Two tool block stops (at indices 1 and 2), plus text block stop at 0
+        block_stops = [e for e in parsed if e.get("type") == "content_block_stop"]
+        stop_indices = [e["index"] for e in block_stops]
+        assert 0 in stop_indices
+        assert 1 in stop_indices
+        assert 2 in stop_indices

src/exo/master/tests/test_master.py

@@ -7,15 +7,14 @@ from loguru import logger
 
 from exo.master.main import Master
 from exo.routing.router import get_node_id_keypair
-from exo.shared.models.model_cards import ModelCard, ModelId, ModelTask
-from exo.shared.types.api import ChatCompletionMessage, ChatCompletionTaskParams
+from exo.shared.models.model_cards import ModelCard, ModelTask
 from exo.shared.types.commands import (
-    ChatCompletion,
     CommandId,
     ForwarderCommand,
     PlaceInstance,
+    TextGeneration,
 )
-from exo.shared.types.common import NodeId, SessionId
+from exo.shared.types.common import ModelId, NodeId, SessionId
 from exo.shared.types.events import (
     ForwarderEvent,
     IndexedEvent,
@@ -27,8 +26,9 @@ from exo.shared.types.memory import Memory
 from exo.shared.types.profiling import (
     MemoryUsage,
 )
-from exo.shared.types.tasks import ChatCompletion as ChatCompletionTask
 from exo.shared.types.tasks import TaskStatus
+from exo.shared.types.tasks import TextGeneration as TextGenerationTask
+from exo.shared.types.text_generation import InputMessage, TextGenerationTaskParams
 from exo.shared.types.worker.instances import (
     InstanceMeta,
     MlxRingInstance,
@@ -127,19 +127,17 @@ async def test_master():
         logger.info("wait for an instance")
         while len(master.state.instances.keys()) == 0:
             await anyio.sleep(0.001)
-        logger.info("inject a ChatCompletion Command")
+        logger.info("inject a TextGeneration Command")
         await command_sender.send(
             ForwarderCommand(
                 origin=node_id,
                 command=(
-                    ChatCompletion(
+                    TextGeneration(
                         command_id=CommandId(),
-                        request_params=ChatCompletionTaskParams(
-                            model="llama-3.2-1b",
-                            messages=[
-                                ChatCompletionMessage(
-                                    role="user", content="Hello, how are you?"
-                                )
+                        task_params=TextGenerationTaskParams(
+                            model=ModelId("llama-3.2-1b"),
+                            input=[
+                                InputMessage(role="user", content="Hello, how are you?")
                             ],
                         ),
                     )
@@ -190,12 +188,10 @@ async def test_master():
         assert created_instance.ephemeral_port > 0
         assert isinstance(events[2].event, TaskCreated)
         assert events[2].event.task.task_status == TaskStatus.Pending
-        assert isinstance(events[2].event.task, ChatCompletionTask)
-        assert events[2].event.task.task_params == ChatCompletionTaskParams(
-            model="llama-3.2-1b",
-            messages=[
-                ChatCompletionMessage(role="user", content="Hello, how are you?")
-            ],
+        assert isinstance(events[2].event.task, TextGenerationTask)
+        assert events[2].event.task.task_params == TextGenerationTaskParams(
+            model=ModelId("llama-3.2-1b"),
+            input=[InputMessage(role="user", content="Hello, how are you?")],
         )
 
         await master.shutdown()

src/exo/master/tests/test_openai_responses_api.py

@@ -0,0 +1,48 @@
+"""Tests for OpenAI Responses API wire types.
+
+ResponsesRequest is the API wire type for the Responses endpoint.
+The responses adapter converts it to TextGenerationTaskParams for the pipeline.
+"""
+
+import pydantic
+import pytest
+
+from exo.shared.types.common import ModelId
+from exo.shared.types.openai_responses import (
+    ResponseInputMessage,
+    ResponsesRequest,
+)
+
+
+class TestResponsesRequestValidation:
+    """Tests for OpenAI Responses API request validation."""
+
+    def test_request_requires_model(self):
+        with pytest.raises(pydantic.ValidationError):
+            ResponsesRequest.model_validate(
+                {
+                    "input": "Hello",
+                }
+            )
+
+    def test_request_requires_input(self):
+        with pytest.raises(pydantic.ValidationError):
+            ResponsesRequest.model_validate(
+                {
+                    "model": "gpt-4o",
+                }
+            )
+
+    def test_request_accepts_string_input(self):
+        request = ResponsesRequest(
+            model=ModelId("gpt-4o"),
+            input="Hello",
+        )
+        assert request.input == "Hello"
+
+    def test_request_accepts_message_array_input(self):
+        request = ResponsesRequest(
+            model=ModelId("gpt-4o"),
+            input=[ResponseInputMessage(role="user", content="Hello")],
+        )
+        assert len(request.input) == 1

src/exo/master/tests/test_placement_utils.py

@@ -5,6 +5,7 @@ from exo.master.placement_utils import (
     filter_cycles_by_memory,
     get_mlx_jaccl_coordinators,
     get_shard_assignments,
+    get_shard_assignments_for_pipeline_parallel,
     get_smallest_cycles,
 )
 from exo.master.tests.conftest import (
@@ -20,7 +21,11 @@ from exo.shared.types.profiling import (
     NodeNetworkInfo,
 )
 from exo.shared.types.topology import Connection, SocketConnection
-from exo.shared.types.worker.shards import Sharding
+from exo.shared.types.worker.shards import (
+    CfgShardMetadata,
+    PipelineShardMetadata,
+    Sharding,
+)
 
 
 def test_filter_cycles_by_memory():
@@ -487,3 +492,193 @@ def test_get_shard_assignments_insufficient_memory_raises():
         get_shard_assignments(
             model_card, selected_cycle, Sharding.Pipeline, node_memory
         )
+
+
+class TestCfgParallelPlacement:
+    def _create_ring_topology(self, node_ids: list[NodeId]) -> Topology:
+        topology = Topology()
+        for node_id in node_ids:
+            topology.add_node(node_id)
+
+        for i, node_id in enumerate(node_ids):
+            next_node = node_ids[(i + 1) % len(node_ids)]
+            conn = Connection(
+                source=node_id,
+                sink=next_node,
+                edge=create_socket_connection(i + 1),
+            )
+            topology.add_connection(conn)
+
+        return topology
+
+    def test_two_nodes_cfg_model_uses_cfg_parallel(self):
+        """Two nodes with CFG model should use CFG parallel (no pipeline)."""
+        node_a = NodeId()
+        node_b = NodeId()
+
+        topology = self._create_ring_topology([node_a, node_b])
+        cycles = [c for c in topology.get_cycles() if len(c) == 2]
+        cycle = cycles[0]
+
+        node_memory = {
+            node_a: create_node_memory(1000 * 1024),
+            node_b: create_node_memory(1000 * 1024),
+        }
+
+        model_card = ModelCard(
+            model_id=ModelId("qwen-image-test"),
+            n_layers=60,
+            storage_size=Memory.from_kb(1000),
+            hidden_size=1,
+            supports_tensor=False,
+            uses_cfg=True,
+            tasks=[ModelTask.TextToImage],
+        )
+
+        assignments = get_shard_assignments_for_pipeline_parallel(
+            model_card, cycle, node_memory
+        )
+
+        shards = list(assignments.runner_to_shard.values())
+        assert len(shards) == 2
+
+        # CFG models should get CfgShardMetadata
+        for shard in shards:
+            assert isinstance(shard, CfgShardMetadata)
+            # Both nodes should have all layers (no pipeline split)
+            assert shard.start_layer == 0
+            assert shard.end_layer == 60
+            assert shard.cfg_world_size == 2
+            # Each node is the only stage in its pipeline group
+            assert shard.pipeline_world_size == 1
+            assert shard.pipeline_rank == 0
+
+        cfg_ranks = sorted(
+            s.cfg_rank for s in shards if isinstance(s, CfgShardMetadata)
+        )
+        assert cfg_ranks == [0, 1]
+
+    def test_four_nodes_cfg_model_uses_hybrid(self):
+        """Four nodes with CFG model should use 2 CFG groups x 2 pipeline stages."""
+        nodes = [NodeId() for _ in range(4)]
+
+        topology = self._create_ring_topology(nodes)
+        cycles = [c for c in topology.get_cycles() if len(c) == 4]
+        cycle = cycles[0]
+
+        node_memory = {n: create_node_memory(1000 * 1024) for n in nodes}
+
+        model_card = ModelCard(
+            model_id=ModelId("qwen-image-test"),
+            n_layers=60,
+            storage_size=Memory.from_kb(1000),
+            hidden_size=1,
+            supports_tensor=False,
+            uses_cfg=True,
+            tasks=[ModelTask.TextToImage],
+        )
+
+        assignments = get_shard_assignments_for_pipeline_parallel(
+            model_card, cycle, node_memory
+        )
+
+        shards = list(assignments.runner_to_shard.values())
+        assert len(shards) == 4
+
+        # CFG models should get CfgShardMetadata
+        for shard in shards:
+            assert isinstance(shard, CfgShardMetadata)
+            assert shard.cfg_world_size == 2
+            assert shard.pipeline_world_size == 2
+            assert shard.pipeline_rank in [0, 1]
+
+        # Check we have 2 nodes in each CFG group
+        cfg_0_shards = [
+            s for s in shards if isinstance(s, CfgShardMetadata) and s.cfg_rank == 0
+        ]
+        cfg_1_shards = [
+            s for s in shards if isinstance(s, CfgShardMetadata) and s.cfg_rank == 1
+        ]
+        assert len(cfg_0_shards) == 2
+        assert len(cfg_1_shards) == 2
+
+        # Both CFG groups should have the same layer assignments
+        cfg_0_layers = [(s.start_layer, s.end_layer) for s in cfg_0_shards]
+        cfg_1_layers = [(s.start_layer, s.end_layer) for s in cfg_1_shards]
+        assert sorted(cfg_0_layers) == sorted(cfg_1_layers)
+
+    def test_three_nodes_cfg_model_uses_sequential_cfg(self):
+        """Three nodes (odd) with CFG model should use sequential CFG (PipelineShardMetadata)."""
+        nodes = [NodeId() for _ in range(3)]
+
+        topology = self._create_ring_topology(nodes)
+        cycles = [c for c in topology.get_cycles() if len(c) == 3]
+        cycle = cycles[0]
+
+        node_memory = {n: create_node_memory(1000 * 1024) for n in nodes}
+
+        model_card = ModelCard(
+            model_id=ModelId("qwen-image-test"),
+            n_layers=60,
+            storage_size=Memory.from_kb(1000),
+            hidden_size=1,
+            supports_tensor=False,
+            uses_cfg=True,
+            tasks=[ModelTask.TextToImage],
+        )
+
+        assignments = get_shard_assignments_for_pipeline_parallel(
+            model_card, cycle, node_memory
+        )
+
+        shards = list(assignments.runner_to_shard.values())
+        assert len(shards) == 3
+
+        # Odd node count with CFG model falls back to PipelineShardMetadata (sequential CFG)
+        for shard in shards:
+            assert isinstance(shard, PipelineShardMetadata)
+
+    def test_two_nodes_non_cfg_model_uses_pipeline(self):
+        """Two nodes with non-CFG model should use pure pipeline (PipelineShardMetadata)."""
+        node_a = NodeId()
+        node_b = NodeId()
+
+        topology = self._create_ring_topology([node_a, node_b])
+        cycles = [c for c in topology.get_cycles() if len(c) == 2]
+        cycle = cycles[0]
+
+        node_memory = {
+            node_a: create_node_memory(1000 * 1024),
+            node_b: create_node_memory(1000 * 1024),
+        }
+
+        model_card = ModelCard(
+            model_id=ModelId("flux-test"),
+            n_layers=57,
+            storage_size=Memory.from_kb(1000),
+            hidden_size=1,
+            supports_tensor=False,
+            uses_cfg=False,  # Non-CFG model
+            tasks=[ModelTask.TextToImage],
+        )
+
+        assignments = get_shard_assignments_for_pipeline_parallel(
+            model_card, cycle, node_memory
+        )
+
+        shards = list(assignments.runner_to_shard.values())
+        assert len(shards) == 2
+
+        # Non-CFG models should get PipelineShardMetadata
+        for shard in shards:
+            assert isinstance(shard, PipelineShardMetadata)
+
+        # Should have actual layer sharding (pipeline)
+        layer_ranges = sorted(
+            (s.start_layer, s.end_layer)
+            for s in shards
+            if isinstance(s, PipelineShardMetadata)
+        )
+        # First shard starts at 0, last shard ends at 57
+        assert layer_ranges[0][0] == 0
+        assert layer_ranges[-1][1] == 57

src/exo/shared/apply.py

@@ -25,6 +25,8 @@ from exo.shared.types.events import (
     TestEvent,
     TopologyEdgeCreated,
     TopologyEdgeDeleted,
+    TracesCollected,
+    TracesMerged,
 )
 from exo.shared.types.profiling import (
     NodeIdentity,
@@ -55,7 +57,12 @@ def event_apply(event: Event, state: State) -> State:
     """Apply an event to state."""
     match event:
         case (
-            TestEvent() | ChunkGenerated() | TaskAcknowledged() | InputChunkReceived()
+            TestEvent()
+            | ChunkGenerated()
+            | TaskAcknowledged()
+            | InputChunkReceived()
+            | TracesCollected()
+            | TracesMerged()
         ):  # Pass-through events that don't modify state
             return state
         case InstanceCreated():

src/exo/shared/constants.py

@@ -2,6 +2,8 @@ import os
 import sys
 from pathlib import Path
 
+from exo.utils.dashboard_path import find_dashboard, find_resources
+
 _EXO_HOME_ENV = os.environ.get("EXO_HOME", None)
 
 
@@ -31,6 +33,14 @@ EXO_MODELS_DIR = (
     if _EXO_MODELS_DIR_ENV is None
     else Path.home() / _EXO_MODELS_DIR_ENV
 )
+_RESOURCES_DIR_ENV = os.environ.get("EXO_RESOURCES_DIR", None)
+RESOURCES_DIR = (
+    find_resources() if _RESOURCES_DIR_ENV is None else Path.home() / _RESOURCES_DIR_ENV
+)
+_DASHBOARD_DIR_ENV = os.environ.get("EXO_DASHBOARD_DIR", None)
+DASHBOARD_DIR = (
+    find_dashboard() if _DASHBOARD_DIR_ENV is None else Path.home() / _DASHBOARD_DIR_ENV
+)
 
 # Log files (data/logs or cache)
 EXO_LOG = EXO_CACHE_HOME / "exo.log"
@@ -48,8 +58,13 @@ LIBP2P_COMMANDS_TOPIC = "commands"
 
 EXO_MAX_CHUNK_SIZE = 512 * 1024
 
+EXO_CUSTOM_MODEL_CARDS_DIR = EXO_DATA_HOME / "custom_model_cards"
+
 EXO_IMAGE_CACHE_DIR = EXO_CACHE_HOME / "images"
+EXO_TRACING_CACHE_DIR = EXO_CACHE_HOME / "traces"
 
 EXO_ENABLE_IMAGE_MODELS = (
     os.getenv("EXO_ENABLE_IMAGE_MODELS", "false").lower() == "true"
 )
+
+EXO_TRACING_ENABLED = os.getenv("EXO_TRACING_ENABLED", "false").lower() == "true"

src/exo/shared/models/model_cards.py

@@ -12,16 +12,47 @@ from pydantic import (
     BaseModel,
     Field,
     PositiveInt,
+    ValidationError,
     field_validator,
     model_validator,
 )
+from tomlkit.exceptions import TOMLKitError
 
-from exo.shared.constants import EXO_ENABLE_IMAGE_MODELS
+from exo.shared.constants import (
+    EXO_CUSTOM_MODEL_CARDS_DIR,
+    EXO_ENABLE_IMAGE_MODELS,
+    RESOURCES_DIR,
+)
 from exo.shared.types.common import ModelId
 from exo.shared.types.memory import Memory
 from exo.utils.pydantic_ext import CamelCaseModel
 
-_card_cache: dict[str, "ModelCard"] = {}
+# kinda ugly...
+# TODO: load search path from config.toml
+_custom_cards_dir = Path(str(EXO_CUSTOM_MODEL_CARDS_DIR))
+_csp = [Path(RESOURCES_DIR) / "inference_model_cards", _custom_cards_dir]
+if EXO_ENABLE_IMAGE_MODELS:
+    _csp.append(Path(RESOURCES_DIR) / "image_model_cards")
+
+CARD_SEARCH_PATH = _csp
+
+_card_cache: dict[ModelId, "ModelCard"] = {}
+
+
+async def _refresh_card_cache():
+    for path in CARD_SEARCH_PATH:
+        async for toml_file in path.rglob("*.toml"):
+            try:
+                card = await ModelCard.load_from_path(toml_file)
+                _card_cache[card.model_id] = card
+            except (ValidationError, TOMLKitError):
+                pass
+
+
+async def get_model_cards() -> list["ModelCard"]:
+    if len(_card_cache) == 0:
+        await _refresh_card_cache()
+    return list(_card_cache.values())
 
 
 class ModelTask(str, Enum):
@@ -34,9 +65,9 @@ class ComponentInfo(CamelCaseModel):
     component_name: str
     component_path: str
     storage_size: Memory
-    n_layers: PositiveInt | None
+    n_layers: PositiveInt | None = None
     can_shard: bool
-    safetensors_index_filename: str | None
+    safetensors_index_filename: str | None = None
 
 
 class ModelCard(CamelCaseModel):
@@ -47,6 +78,11 @@ class ModelCard(CamelCaseModel):
     supports_tensor: bool
     tasks: list[ModelTask]
     components: list[ComponentInfo] | None = None
+    family: str = ""
+    quantization: str = ""
+    base_model: str = ""
+    capabilities: list[str] = []
+    uses_cfg: bool = False
 
     @field_validator("tasks", mode="before")
     @classmethod
@@ -55,31 +91,37 @@ class ModelCard(CamelCaseModel):
 
     async def save(self, path: Path) -> None:
         async with await open_file(path, "w") as f:
-            py = self.model_dump()
+            py = self.model_dump(exclude_none=True)
             data = tomlkit.dumps(py)  # pyright: ignore[reportUnknownMemberType]
             await f.write(data)
 
+    async def save_to_custom_dir(self) -> None:
+        await aios.makedirs(str(_custom_cards_dir), exist_ok=True)
+        await self.save(_custom_cards_dir / (self.model_id.normalize() + ".toml"))
+
     @staticmethod
     async def load_from_path(path: Path) -> "ModelCard":
         async with await open_file(path, "r") as f:
             py = tomlkit.loads(await f.read())
             return ModelCard.model_validate(py)
 
+    # Is it okay that model card.load defaults to network access if the card doesn't exist? do we want to be more explicit here?
     @staticmethod
     async def load(model_id: ModelId) -> "ModelCard":
-        for card in MODEL_CARDS.values():
-            if card.model_id == model_id:
-                return card
-        return await ModelCard.from_hf(model_id)
-
-    @staticmethod
-    async def from_hf(model_id: ModelId) -> "ModelCard":
-        """Fetches storage size and number of layers for a Hugging Face model, returns Pydantic ModelMeta."""
+        if model_id not in _card_cache:
+            await _refresh_card_cache()
         if (mc := _card_cache.get(model_id)) is not None:
             return mc
-        config_data = await get_config_data(model_id)
+
+        return await ModelCard.fetch_from_hf(model_id)
+
+    @staticmethod
+    async def fetch_from_hf(model_id: ModelId) -> "ModelCard":
+        """Fetches storage size and number of layers for a Hugging Face model, returns Pydantic ModelMeta."""
+        # TODO: failure if files do not exist
+        config_data = await fetch_config_data(model_id)
         num_layers = config_data.layer_count
-        mem_size_bytes = await get_safetensors_size(model_id)
+        mem_size_bytes = await fetch_safetensors_size(model_id)
 
         mc = ModelCard(
             model_id=ModelId(model_id),
@@ -89,630 +131,29 @@ class ModelCard(CamelCaseModel):
             supports_tensor=config_data.supports_tensor,
             tasks=[ModelTask.TextGeneration],
         )
+        await mc.save_to_custom_dir()
         _card_cache[model_id] = mc
         return mc
 
 
-MODEL_CARDS: dict[str, ModelCard] = {
-    # deepseek v3
-    "deepseek-v3.1-4bit": ModelCard(
-        model_id=ModelId("mlx-community/DeepSeek-V3.1-4bit"),
-        storage_size=Memory.from_gb(378),
-        n_layers=61,
-        hidden_size=7168,
-        supports_tensor=True,
-        tasks=[ModelTask.TextGeneration],
-    ),
-    "deepseek-v3.1-8bit": ModelCard(
-        model_id=ModelId("mlx-community/DeepSeek-V3.1-8bit"),
-        storage_size=Memory.from_gb(713),
-        n_layers=61,
-        hidden_size=7168,
-        supports_tensor=True,
-        tasks=[ModelTask.TextGeneration],
-    ),
-    # kimi k2
-    "kimi-k2-instruct-4bit": ModelCard(
-        model_id=ModelId("mlx-community/Kimi-K2-Instruct-4bit"),
-        storage_size=Memory.from_gb(578),
-        n_layers=61,
-        hidden_size=7168,
-        supports_tensor=True,
-        tasks=[ModelTask.TextGeneration],
-    ),
-    "kimi-k2-thinking": ModelCard(
-        model_id=ModelId("mlx-community/Kimi-K2-Thinking"),
-        storage_size=Memory.from_gb(658),
-        n_layers=61,
-        hidden_size=7168,
-        supports_tensor=True,
-        tasks=[ModelTask.TextGeneration],
-    ),
-    "kimi-k2.5": ModelCard(
-        model_id=ModelId("mlx-community/Kimi-K2.5"),
-        storage_size=Memory.from_gb(617),
-        n_layers=61,
-        hidden_size=7168,
-        supports_tensor=True,
-        tasks=[ModelTask.TextGeneration],
-    ),
-    # llama-3.1
-    "llama-3.1-8b": ModelCard(
-        model_id=ModelId("mlx-community/Meta-Llama-3.1-8B-Instruct-4bit"),
-        storage_size=Memory.from_mb(4423),
-        n_layers=32,
-        hidden_size=4096,
-        supports_tensor=True,
-        tasks=[ModelTask.TextGeneration],
-    ),
-    "llama-3.1-8b-8bit": ModelCard(
-        model_id=ModelId("mlx-community/Meta-Llama-3.1-8B-Instruct-8bit"),
-        storage_size=Memory.from_mb(8540),
-        n_layers=32,
-        hidden_size=4096,
-        supports_tensor=True,
-        tasks=[ModelTask.TextGeneration],
-    ),
-    "llama-3.1-8b-bf16": ModelCard(
-        model_id=ModelId("mlx-community/Meta-Llama-3.1-8B-Instruct-bf16"),
-        storage_size=Memory.from_mb(16100),
-        n_layers=32,
-        hidden_size=4096,
-        supports_tensor=True,
-        tasks=[ModelTask.TextGeneration],
-    ),
-    "llama-3.1-70b": ModelCard(
-        model_id=ModelId("mlx-community/Meta-Llama-3.1-70B-Instruct-4bit"),
-        storage_size=Memory.from_mb(38769),
-        n_layers=80,
-        hidden_size=8192,
-        supports_tensor=True,
-        tasks=[ModelTask.TextGeneration],
-    ),
-    # llama-3.2
-    "llama-3.2-1b": ModelCard(
-        model_id=ModelId("mlx-community/Llama-3.2-1B-Instruct-4bit"),
-        storage_size=Memory.from_mb(696),
-        n_layers=16,
-        hidden_size=2048,
-        supports_tensor=True,
-        tasks=[ModelTask.TextGeneration],
-    ),
-    "llama-3.2-3b": ModelCard(
-        model_id=ModelId("mlx-community/Llama-3.2-3B-Instruct-4bit"),
-        storage_size=Memory.from_mb(1777),
-        n_layers=28,
-        hidden_size=3072,
-        supports_tensor=True,
-        tasks=[ModelTask.TextGeneration],
-    ),
-    "llama-3.2-3b-8bit": ModelCard(
-        model_id=ModelId("mlx-community/Llama-3.2-3B-Instruct-8bit"),
-        storage_size=Memory.from_mb(3339),
-        n_layers=28,
-        hidden_size=3072,
-        supports_tensor=True,
-        tasks=[ModelTask.TextGeneration],
-    ),
-    # llama-3.3
-    "llama-3.3-70b": ModelCard(
-        model_id=ModelId("mlx-community/Llama-3.3-70B-Instruct-4bit"),
-        storage_size=Memory.from_mb(38769),
-        n_layers=80,
-        hidden_size=8192,
-        supports_tensor=True,
-        tasks=[ModelTask.TextGeneration],
-    ),
-    "llama-3.3-70b-8bit": ModelCard(
-        model_id=ModelId("mlx-community/Llama-3.3-70B-Instruct-8bit"),
-        storage_size=Memory.from_mb(73242),
-        n_layers=80,
-        hidden_size=8192,
-        supports_tensor=True,
-        tasks=[ModelTask.TextGeneration],
-    ),
-    "llama-3.3-70b-fp16": ModelCard(
-        model_id=ModelId("mlx-community/llama-3.3-70b-instruct-fp16"),
-        storage_size=Memory.from_mb(137695),
-        n_layers=80,
-        hidden_size=8192,
-        supports_tensor=True,
-        tasks=[ModelTask.TextGeneration],
-    ),
-    # qwen3
-    "qwen3-0.6b": ModelCard(
-        model_id=ModelId("mlx-community/Qwen3-0.6B-4bit"),
-        storage_size=Memory.from_mb(327),
-        n_layers=28,
-        hidden_size=1024,
-        supports_tensor=False,
-        tasks=[ModelTask.TextGeneration],
-    ),
-    "qwen3-0.6b-8bit": ModelCard(
-        model_id=ModelId("mlx-community/Qwen3-0.6B-8bit"),
-        storage_size=Memory.from_mb(666),
-        n_layers=28,
-        hidden_size=1024,
-        supports_tensor=False,
-        tasks=[ModelTask.TextGeneration],
-    ),
-    "qwen3-30b": ModelCard(
-        model_id=ModelId("mlx-community/Qwen3-30B-A3B-4bit"),
-        storage_size=Memory.from_mb(16797),
-        n_layers=48,
-        hidden_size=2048,
-        supports_tensor=True,
-        tasks=[ModelTask.TextGeneration],
-    ),
-    "qwen3-30b-8bit": ModelCard(
-        model_id=ModelId("mlx-community/Qwen3-30B-A3B-8bit"),
-        storage_size=Memory.from_mb(31738),
-        n_layers=48,
-        hidden_size=2048,
-        supports_tensor=True,
-        tasks=[ModelTask.TextGeneration],
-    ),
-    "qwen3-80b-a3B-4bit": ModelCard(
-        model_id=ModelId("mlx-community/Qwen3-Next-80B-A3B-Instruct-4bit"),
-        storage_size=Memory.from_mb(44800),
-        n_layers=48,
-        hidden_size=2048,
-        supports_tensor=True,
-        tasks=[ModelTask.TextGeneration],
-    ),
-    "qwen3-80b-a3B-8bit": ModelCard(
-        model_id=ModelId("mlx-community/Qwen3-Next-80B-A3B-Instruct-8bit"),
-        storage_size=Memory.from_mb(84700),
-        n_layers=48,
-        hidden_size=2048,
-        supports_tensor=True,
-        tasks=[ModelTask.TextGeneration],
-    ),
-    "qwen3-80b-a3B-thinking-4bit": ModelCard(
-        model_id=ModelId("mlx-community/Qwen3-Next-80B-A3B-Thinking-4bit"),
-        storage_size=Memory.from_mb(84700),
-        n_layers=48,
-        hidden_size=2048,
-        supports_tensor=True,
-        tasks=[ModelTask.TextGeneration],
-    ),
-    "qwen3-80b-a3B-thinking-8bit": ModelCard(
-        model_id=ModelId("mlx-community/Qwen3-Next-80B-A3B-Thinking-8bit"),
-        storage_size=Memory.from_mb(84700),
-        n_layers=48,
-        hidden_size=2048,
-        supports_tensor=True,
-        tasks=[ModelTask.TextGeneration],
-    ),
-    "qwen3-235b-a22b-4bit": ModelCard(
-        model_id=ModelId("mlx-community/Qwen3-235B-A22B-Instruct-2507-4bit"),
-        storage_size=Memory.from_gb(132),
-        n_layers=94,
-        hidden_size=4096,
-        supports_tensor=True,
-        tasks=[ModelTask.TextGeneration],
-    ),
-    "qwen3-235b-a22b-8bit": ModelCard(
-        model_id=ModelId("mlx-community/Qwen3-235B-A22B-Instruct-2507-8bit"),
-        storage_size=Memory.from_gb(250),
-        n_layers=94,
-        hidden_size=4096,
-        supports_tensor=True,
-        tasks=[ModelTask.TextGeneration],
-    ),
-    "qwen3-coder-480b-a35b-4bit": ModelCard(
-        model_id=ModelId("mlx-community/Qwen3-Coder-480B-A35B-Instruct-4bit"),
-        storage_size=Memory.from_gb(270),
-        n_layers=62,
-        hidden_size=6144,
-        supports_tensor=True,
-        tasks=[ModelTask.TextGeneration],
-    ),
-    "qwen3-coder-480b-a35b-8bit": ModelCard(
-        model_id=ModelId("mlx-community/Qwen3-Coder-480B-A35B-Instruct-8bit"),
-        storage_size=Memory.from_gb(540),
-        n_layers=62,
-        hidden_size=6144,
-        supports_tensor=True,
-        tasks=[ModelTask.TextGeneration],
-    ),
-    # gpt-oss
-    "gpt-oss-120b-MXFP4-Q8": ModelCard(
-        model_id=ModelId("mlx-community/gpt-oss-120b-MXFP4-Q8"),
-        storage_size=Memory.from_kb(68_996_301),
-        n_layers=36,
-        hidden_size=2880,
-        supports_tensor=True,
-        tasks=[ModelTask.TextGeneration],
-    ),
-    "gpt-oss-20b-MXFP4-Q8": ModelCard(
-        model_id=ModelId("mlx-community/gpt-oss-20b-MXFP4-Q8"),
-        storage_size=Memory.from_kb(11_744_051),
-        n_layers=24,
-        hidden_size=2880,
-        supports_tensor=True,
-        tasks=[ModelTask.TextGeneration],
-    ),
-    # glm 4.5
-    "glm-4.5-air-8bit": ModelCard(
-        # Needs to be quantized g32 or g16 to work with tensor parallel
-        model_id=ModelId("mlx-community/GLM-4.5-Air-8bit"),
-        storage_size=Memory.from_gb(114),
-        n_layers=46,
-        hidden_size=4096,
-        supports_tensor=False,
-        tasks=[ModelTask.TextGeneration],
-    ),
-    "glm-4.5-air-bf16": ModelCard(
-        model_id=ModelId("mlx-community/GLM-4.5-Air-bf16"),
-        storage_size=Memory.from_gb(214),
-        n_layers=46,
-        hidden_size=4096,
-        supports_tensor=True,
-        tasks=[ModelTask.TextGeneration],
-    ),
-    # glm 4.7
-    "glm-4.7-4bit": ModelCard(
-        model_id=ModelId("mlx-community/GLM-4.7-4bit"),
-        storage_size=Memory.from_bytes(198556925568),
-        n_layers=91,
-        hidden_size=5120,
-        supports_tensor=True,
-        tasks=[ModelTask.TextGeneration],
-    ),
-    "glm-4.7-6bit": ModelCard(
-        model_id=ModelId("mlx-community/GLM-4.7-6bit"),
-        storage_size=Memory.from_bytes(286737579648),
-        n_layers=91,
-        hidden_size=5120,
-        supports_tensor=True,
-        tasks=[ModelTask.TextGeneration],
-    ),
-    "glm-4.7-8bit-gs32": ModelCard(
-        model_id=ModelId("mlx-community/GLM-4.7-8bit-gs32"),
-        storage_size=Memory.from_bytes(396963397248),
-        n_layers=91,
-        hidden_size=5120,
-        supports_tensor=True,
-        tasks=[ModelTask.TextGeneration],
-    ),
-    # glm 4.7 flash
-    "glm-4.7-flash-4bit": ModelCard(
-        model_id=ModelId("mlx-community/GLM-4.7-Flash-4bit"),
-        storage_size=Memory.from_gb(18),
-        n_layers=47,
-        hidden_size=2048,
-        supports_tensor=True,
-        tasks=[ModelTask.TextGeneration],
-    ),
-    "glm-4.7-flash-5bit": ModelCard(
-        model_id=ModelId("mlx-community/GLM-4.7-Flash-5bit"),
-        storage_size=Memory.from_gb(21),
-        n_layers=47,
-        hidden_size=2048,
-        supports_tensor=True,
-        tasks=[ModelTask.TextGeneration],
-    ),
-    "glm-4.7-flash-6bit": ModelCard(
-        model_id=ModelId("mlx-community/GLM-4.7-Flash-6bit"),
-        storage_size=Memory.from_gb(25),
-        n_layers=47,
-        hidden_size=2048,
-        supports_tensor=True,
-        tasks=[ModelTask.TextGeneration],
-    ),
-    "glm-4.7-flash-8bit": ModelCard(
-        model_id=ModelId("mlx-community/GLM-4.7-Flash-8bit"),
-        storage_size=Memory.from_gb(32),
-        n_layers=47,
-        hidden_size=2048,
-        supports_tensor=True,
-        tasks=[ModelTask.TextGeneration],
-    ),
-    # minimax-m2
-    "minimax-m2.1-8bit": ModelCard(
-        model_id=ModelId("mlx-community/MiniMax-M2.1-8bit"),
-        storage_size=Memory.from_bytes(242986745856),
-        n_layers=61,
-        hidden_size=3072,
-        supports_tensor=True,
-        tasks=[ModelTask.TextGeneration],
-    ),
-    "minimax-m2.1-3bit": ModelCard(
-        model_id=ModelId("mlx-community/MiniMax-M2.1-3bit"),
-        storage_size=Memory.from_bytes(100086644736),
-        n_layers=61,
-        hidden_size=3072,
-        supports_tensor=True,
-        tasks=[ModelTask.TextGeneration],
-    ),
-}
-
-_IMAGE_BASE_MODEL_CARDS: dict[str, ModelCard] = {
-    "flux1-schnell": ModelCard(
-        model_id=ModelId("exolabs/FLUX.1-schnell"),
-        storage_size=Memory.from_bytes(23782357120 + 9524621312),
-        n_layers=57,
-        hidden_size=1,
-        supports_tensor=False,
-        tasks=[ModelTask.TextToImage],
-        components=[
-            ComponentInfo(
-                component_name="text_encoder",
-                component_path="text_encoder/",
-                storage_size=Memory.from_kb(0),
-                n_layers=12,
-                can_shard=False,
-                safetensors_index_filename=None,
-            ),
-            ComponentInfo(
-                component_name="text_encoder_2",
-                component_path="text_encoder_2/",
-                storage_size=Memory.from_bytes(9524621312),
-                n_layers=24,
-                can_shard=False,
-                safetensors_index_filename="model.safetensors.index.json",
-            ),
-            ComponentInfo(
-                component_name="transformer",
-                component_path="transformer/",
-                storage_size=Memory.from_bytes(23782357120),
-                n_layers=57,
-                can_shard=True,
-                safetensors_index_filename="diffusion_pytorch_model.safetensors.index.json",
-            ),
-            ComponentInfo(
-                component_name="vae",
-                component_path="vae/",
-                storage_size=Memory.from_kb(0),
-                n_layers=None,
-                can_shard=False,
-                safetensors_index_filename=None,
-            ),
-        ],
-    ),
-    "flux1-dev": ModelCard(
-        model_id=ModelId("exolabs/FLUX.1-dev"),
-        storage_size=Memory.from_bytes(23782357120 + 9524621312),
-        n_layers=57,
-        hidden_size=1,
-        supports_tensor=False,
-        tasks=[ModelTask.TextToImage],
-        components=[
-            ComponentInfo(
-                component_name="text_encoder",
-                component_path="text_encoder/",
-                storage_size=Memory.from_kb(0),
-                n_layers=12,
-                can_shard=False,
-                safetensors_index_filename=None,
-            ),
-            ComponentInfo(
-                component_name="text_encoder_2",
-                component_path="text_encoder_2/",
-                storage_size=Memory.from_bytes(9524621312),
-                n_layers=24,
-                can_shard=False,
-                safetensors_index_filename="model.safetensors.index.json",
-            ),
-            ComponentInfo(
-                component_name="transformer",
-                component_path="transformer/",
-                storage_size=Memory.from_bytes(23802816640),
-                n_layers=57,
-                can_shard=True,
-                safetensors_index_filename="diffusion_pytorch_model.safetensors.index.json",
-            ),
-            ComponentInfo(
-                component_name="vae",
-                component_path="vae/",
-                storage_size=Memory.from_kb(0),
-                n_layers=None,
-                can_shard=False,
-                safetensors_index_filename=None,
-            ),
-        ],
-    ),
-    "flux1-krea-dev": ModelCard(
-        model_id=ModelId("exolabs/FLUX.1-Krea-dev"),
-        storage_size=Memory.from_bytes(23802816640 + 9524621312),  # Same as dev
-        n_layers=57,
-        hidden_size=1,
-        supports_tensor=False,
-        tasks=[ModelTask.TextToImage],
-        components=[
-            ComponentInfo(
-                component_name="text_encoder",
-                component_path="text_encoder/",
-                storage_size=Memory.from_kb(0),
-                n_layers=12,
-                can_shard=False,
-                safetensors_index_filename=None,
-            ),
-            ComponentInfo(
-                component_name="text_encoder_2",
-                component_path="text_encoder_2/",
-                storage_size=Memory.from_bytes(9524621312),
-                n_layers=24,
-                can_shard=False,
-                safetensors_index_filename="model.safetensors.index.json",
-            ),
-            ComponentInfo(
-                component_name="transformer",
-                component_path="transformer/",
-                storage_size=Memory.from_bytes(23802816640),
-                n_layers=57,
-                can_shard=True,
-                safetensors_index_filename="diffusion_pytorch_model.safetensors.index.json",
-            ),
-            ComponentInfo(
-                component_name="vae",
-                component_path="vae/",
-                storage_size=Memory.from_kb(0),
-                n_layers=None,
-                can_shard=False,
-                safetensors_index_filename=None,
-            ),
-        ],
-    ),
-    "qwen-image": ModelCard(
-        model_id=ModelId("exolabs/Qwen-Image"),
-        storage_size=Memory.from_bytes(16584333312 + 40860802176),
-        n_layers=60,
-        hidden_size=1,
-        supports_tensor=False,
-        tasks=[ModelTask.TextToImage],
-        components=[
-            ComponentInfo(
-                component_name="text_encoder",
-                component_path="text_encoder/",
-                storage_size=Memory.from_bytes(16584333312),
-                n_layers=12,
-                can_shard=False,
-                safetensors_index_filename=None,
-            ),
-            ComponentInfo(
-                component_name="transformer",
-                component_path="transformer/",
-                storage_size=Memory.from_bytes(40860802176),
-                n_layers=60,
-                can_shard=True,
-                safetensors_index_filename="diffusion_pytorch_model.safetensors.index.json",
-            ),
-            ComponentInfo(
-                component_name="vae",
-                component_path="vae/",
-                storage_size=Memory.from_kb(0),
-                n_layers=None,
-                can_shard=False,
-                safetensors_index_filename=None,
-            ),
-        ],
-    ),
-    "qwen-image-edit-2509": ModelCard(
-        model_id=ModelId("exolabs/Qwen-Image-Edit-2509"),
-        storage_size=Memory.from_bytes(16584333312 + 40860802176),
-        n_layers=60,
-        hidden_size=1,
-        supports_tensor=False,
-        tasks=[ModelTask.ImageToImage],
-        components=[
-            ComponentInfo(
-                component_name="text_encoder",
-                component_path="text_encoder/",
-                storage_size=Memory.from_bytes(16584333312),
-                n_layers=12,
-                can_shard=False,
-                safetensors_index_filename=None,
-            ),
-            ComponentInfo(
-                component_name="transformer",
-                component_path="transformer/",
-                storage_size=Memory.from_bytes(40860802176),
-                n_layers=60,
-                can_shard=True,
-                safetensors_index_filename="diffusion_pytorch_model.safetensors.index.json",
-            ),
-            ComponentInfo(
-                component_name="vae",
-                component_path="vae/",
-                storage_size=Memory.from_kb(0),
-                n_layers=None,
-                can_shard=False,
-                safetensors_index_filename=None,
-            ),
-        ],
-    ),
-}
+async def delete_custom_card(model_id: ModelId) -> bool:
+    """Delete a user-added custom model card. Returns True if deleted."""
+    card_path = _custom_cards_dir / (ModelId(model_id).normalize() + ".toml")
+    if await card_path.exists():
+        await card_path.unlink()
+        _card_cache.pop(model_id, None)
+        return True
+    return False
 
 
-def _generate_image_model_quant_variants(
-    base_name: str,
-    base_card: ModelCard,
-) -> dict[str, ModelCard]:
-    """Create quantized variants of an image model card.
+def is_custom_card(model_id: ModelId) -> bool:
+    """Check if a model card exists in the custom cards directory."""
+    import os
 
-    Only the transformer component is quantized; text encoders stay at bf16.
-    Sizes are calculated exactly from the base card's component sizes.
-    """
-    if base_card.components is None:
-        raise ValueError(f"Image model {base_name} must have components defined")
-
-    # quantizations = [8, 6, 5, 4, 3]
-    quantizations = [8, 4]
-
-    num_transformer_bytes = next(
-        c.storage_size.in_bytes
-        for c in base_card.components
-        if c.component_name == "transformer"
+    card_path = Path(str(EXO_CUSTOM_MODEL_CARDS_DIR)) / (
+        ModelId(model_id).normalize() + ".toml"
     )
-
-    transformer_bytes = Memory.from_bytes(num_transformer_bytes)
-
-    remaining_bytes = Memory.from_bytes(
-        sum(
-            c.storage_size.in_bytes
-            for c in base_card.components
-            if c.component_name != "transformer"
-        )
-    )
-
-    def with_transformer_size(new_size: Memory) -> list[ComponentInfo]:
-        assert base_card.components is not None
-        return [
-            ComponentInfo(
-                component_name=c.component_name,
-                component_path=c.component_path,
-                storage_size=new_size
-                if c.component_name == "transformer"
-                else c.storage_size,
-                n_layers=c.n_layers,
-                can_shard=c.can_shard,
-                safetensors_index_filename=c.safetensors_index_filename,
-            )
-            for c in base_card.components
-        ]
-
-    variants = {
-        base_name: ModelCard(
-            model_id=base_card.model_id,
-            storage_size=transformer_bytes + remaining_bytes,
-            n_layers=base_card.n_layers,
-            hidden_size=base_card.hidden_size,
-            supports_tensor=base_card.supports_tensor,
-            tasks=base_card.tasks,
-            components=with_transformer_size(transformer_bytes),
-        )
-    }
-
-    for quant in quantizations:
-        quant_transformer_bytes = Memory.from_bytes(
-            (num_transformer_bytes * quant) // 16
-        )
-        total_bytes = remaining_bytes + quant_transformer_bytes
-
-        model_id = ModelId(base_card.model_id + f"-{quant}bit")
-
-        variants[f"{base_name}-{quant}bit"] = ModelCard(
-            model_id=model_id,
-            storage_size=total_bytes,
-            n_layers=base_card.n_layers,
-            hidden_size=base_card.hidden_size,
-            supports_tensor=base_card.supports_tensor,
-            tasks=base_card.tasks,
-            components=with_transformer_size(quant_transformer_bytes),
-        )
-
-    return variants
-
-
-_image_model_cards: dict[str, ModelCard] = {}
-for _base_name, _base_card in _IMAGE_BASE_MODEL_CARDS.items():
-    _image_model_cards |= _generate_image_model_quant_variants(_base_name, _base_card)
-_IMAGE_MODEL_CARDS = _image_model_cards
-
-if EXO_ENABLE_IMAGE_MODELS:
-    MODEL_CARDS.update(_IMAGE_MODEL_CARDS)
+    return os.path.isfile(str(card_path))
 
 
 class ConfigData(BaseModel):
@@ -767,7 +208,7 @@ class ConfigData(BaseModel):
         return data
 
 
-async def get_config_data(model_id: ModelId) -> ConfigData:
+async def fetch_config_data(model_id: ModelId) -> ConfigData:
     """Downloads and parses config.json for a model."""
     from exo.download.download_utils import (
         download_file_with_retry,
@@ -789,7 +230,7 @@ async def get_config_data(model_id: ModelId) -> ConfigData:
         return ConfigData.model_validate_json(await f.read())
 
 
-async def get_safetensors_size(model_id: ModelId) -> Memory:
+async def fetch_safetensors_size(model_id: ModelId) -> Memory:
     """Gets model size from safetensors index or falls back to HF API."""
     from exo.download.download_utils import (
         download_file_with_retry,

src/exo/shared/tracing.py

@@ -0,0 +1,238 @@
+from __future__ import annotations
+
+import json
+import time
+from collections import defaultdict
+from collections.abc import Generator
+from contextlib import contextmanager
+from contextvars import ContextVar
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import cast, final
+
+from exo.shared.constants import EXO_TRACING_ENABLED
+from exo.worker.runner.bootstrap import logger
+
+# Context variable to track the current trace category for hierarchical nesting
+_current_category: ContextVar[str | None] = ContextVar("current_category", default=None)
+
+
+@final
+@dataclass(frozen=True)
+class TraceEvent:
+    name: str
+    start_us: int
+    duration_us: int
+    rank: int
+    category: str
+
+
+@final
+@dataclass
+class CategoryStats:
+    total_us: int = 0
+    count: int = 0
+    min_us: int = 0
+    max_us: int = 0
+
+    def add(self, duration_us: int) -> None:
+        if self.count == 0:
+            self.min_us = duration_us
+            self.max_us = duration_us
+        else:
+            self.min_us = min(self.min_us, duration_us)
+            self.max_us = max(self.max_us, duration_us)
+        self.total_us += duration_us
+        self.count += 1
+
+    @property
+    def avg_us(self) -> float:
+        return self.total_us / self.count if self.count > 0 else 0.0
+
+
+@final
+@dataclass
+class TraceStats:
+    total_wall_time_us: int = 0
+    by_category: dict[str, CategoryStats] = field(default_factory=dict)
+    by_rank: dict[int, dict[str, CategoryStats]] = field(default_factory=dict)
+
+
+# Global trace buffer - each rank accumulates traces here
+_trace_buffer: list[TraceEvent] = []
+
+
+def _record_span(
+    name: str, start_us: int, duration_us: int, rank: int, category: str
+) -> None:
+    _trace_buffer.append(
+        TraceEvent(
+            name=name,
+            start_us=start_us,
+            duration_us=duration_us,
+            rank=rank,
+            category=category,
+        )
+    )
+
+
+@contextmanager
+def trace(
+    name: str,
+    rank: int,
+    category: str = "compute",
+) -> Generator[None, None, None]:
+    """Context manager to trace any operation.
+
+    Nested traces automatically inherit the parent category, creating hierarchical
+    categories like "sync/compute" or "async/comms".
+
+    Args:
+        name: Name of the operation (e.g., "recv 0", "send 1", "joint_blocks")
+        rank: This rank's ID
+        category: Category for grouping in trace viewer ("comm", "compute", "step")
+
+    Example:
+        with trace(f"sync {t}", rank, "sync"):
+            with trace("joint_blocks", rank, "compute"):
+                # Recorded with category "sync/compute"
+                hidden_states = some_computation(...)
+    """
+    if not EXO_TRACING_ENABLED:
+        yield
+        return
+
+    # Combine with parent category if nested
+    parent = _current_category.get()
+    full_category = f"{parent}/{category}" if parent else category
+
+    # Set as current for nested traces
+    token = _current_category.set(full_category)
+
+    try:
+        start_us = int(time.time() * 1_000_000)
+        start_perf = time.perf_counter()
+        yield
+        duration_us = int((time.perf_counter() - start_perf) * 1_000_000)
+        _record_span(name, start_us, duration_us, rank, full_category)
+    finally:
+        _current_category.reset(token)
+
+
+def get_trace_buffer() -> list[TraceEvent]:
+    return list(_trace_buffer)
+
+
+def clear_trace_buffer() -> None:
+    _trace_buffer.clear()
+
+
+def export_trace(traces: list[TraceEvent], output_path: Path) -> None:
+    trace_events: list[dict[str, object]] = []
+
+    for event in traces:
+        # Chrome trace format uses "X" for complete events (with duration)
+        chrome_event: dict[str, object] = {
+            "name": event.name,
+            "cat": event.category,
+            "ph": "X",
+            "ts": event.start_us,
+            "dur": event.duration_us,
+            "pid": 0,
+            "tid": event.rank,
+            "args": {"rank": event.rank},
+        }
+        trace_events.append(chrome_event)
+
+    ranks_seen = set(t.rank for t in traces)
+    for rank in ranks_seen:
+        trace_events.append(
+            {
+                "name": "thread_name",
+                "ph": "M",  # Metadata event
+                "pid": 0,
+                "tid": rank,
+                "args": {"name": f"Rank {rank}"},
+            }
+        )
+
+    chrome_trace = {"traceEvents": trace_events}
+
+    try:
+        output_path.parent.mkdir(parents=True, exist_ok=True)
+        with open(output_path, "w") as f:
+            json.dump(chrome_trace, f, indent=2)
+    except OSError as e:
+        logger.warning("Failed to export trace to %s: %s", output_path, e)
+
+
+def load_trace_file(path: Path) -> list[TraceEvent]:
+    with open(path) as f:
+        data = cast(dict[str, list[dict[str, object]]], json.load(f))
+
+    events = data.get("traceEvents", [])
+    traces: list[TraceEvent] = []
+
+    for event in events:
+        # Skip metadata events
+        if event.get("ph") == "M":
+            continue
+
+        name = str(event.get("name", ""))
+        category = str(event.get("cat", ""))
+        ts_value = event.get("ts", 0)
+        dur_value = event.get("dur", 0)
+        tid_value = event.get("tid", 0)
+        start_us = int(ts_value) if isinstance(ts_value, (int, float, str)) else 0
+        duration_us = int(dur_value) if isinstance(dur_value, (int, float, str)) else 0
+
+        # Get rank from tid or args
+        rank = int(tid_value) if isinstance(tid_value, (int, float, str)) else 0
+        args = event.get("args")
+        if isinstance(args, dict):
+            args_dict = cast(dict[str, object], args)
+            rank_from_args = args_dict.get("rank")
+            if isinstance(rank_from_args, (int, float, str)):
+                rank = int(rank_from_args)
+
+        traces.append(
+            TraceEvent(
+                name=name,
+                start_us=start_us,
+                duration_us=duration_us,
+                rank=rank,
+                category=category,
+            )
+        )
+
+    return traces
+
+
+def compute_stats(traces: list[TraceEvent]) -> TraceStats:
+    stats = TraceStats()
+
+    if not traces:
+        return stats
+
+    # Calculate wall time from earliest start to latest end
+    min_start = min(t.start_us for t in traces)
+    max_end = max(t.start_us + t.duration_us for t in traces)
+    stats.total_wall_time_us = max_end - min_start
+
+    # Initialize nested dicts
+    by_category: dict[str, CategoryStats] = defaultdict(CategoryStats)
+    by_rank: dict[int, dict[str, CategoryStats]] = defaultdict(
+        lambda: defaultdict(CategoryStats)
+    )
+
+    for event in traces:
+        # By category
+        by_category[event.category].add(event.duration_us)
+
+        # By rank and category
+        by_rank[event.rank][event.category].add(event.duration_us)
+
+    stats.by_category = dict(by_category)
+    stats.by_rank = {k: dict(v) for k, v in by_rank.items()}
+
+    return stats

src/exo/shared/types/api.py

@@ -2,7 +2,6 @@ import time
 from collections.abc import Generator
 from typing import Annotated, Any, Literal
 
-from fastapi import UploadFile
 from pydantic import BaseModel, Field, field_validator
 from pydantic_core import PydanticUseDefault
 
@@ -11,7 +10,7 @@ from exo.shared.types.common import CommandId, NodeId
 from exo.shared.types.memory import Memory
 from exo.shared.types.worker.instances import Instance, InstanceId, InstanceMeta
 from exo.shared.types.worker.shards import Sharding, ShardMetadata
-from exo.utils.pydantic_ext import CamelCaseModel, ConfigDict, TaggedModel
+from exo.utils.pydantic_ext import CamelCaseModel
 
 FinishReason = Literal[
     "stop", "length", "tool_calls", "content_filter", "function_call", "error"
@@ -43,6 +42,11 @@ class ModelListModel(BaseModel):
     storage_size_megabytes: int = Field(default=0)
     supports_tensor: bool = Field(default=False)
     tasks: list[str] = Field(default=[])
+    is_custom: bool = Field(default=False)
+    family: str = Field(default="")
+    quantization: str = Field(default="")
+    base_model: str = Field(default="")
+    capabilities: list[str] = Field(default_factory=list)
 
 
 class ModelList(BaseModel):
@@ -174,10 +178,8 @@ class StreamOptions(BaseModel):
     include_usage: bool = False
 
 
-class ChatCompletionTaskParams(TaggedModel):
-    model_config = ConfigDict(extra="ignore")
-
-    model: str
+class ChatCompletionRequest(BaseModel):
+    model: ModelId
     frequency_penalty: float | None = None
     messages: list[ChatCompletionMessage]
     logit_bias: dict[str, int] | None = None
@@ -193,16 +195,30 @@ class ChatCompletionTaskParams(TaggedModel):
     stream_options: StreamOptions | None = None
     temperature: float | None = None
     top_p: float | None = None
+    top_k: int | None = None
     tools: list[dict[str, Any]] | None = None
     tool_choice: str | dict[str, Any] | None = None
     parallel_tool_calls: bool | None = None
     user: str | None = None
 
 
-class BenchChatCompletionTaskParams(ChatCompletionTaskParams):
+class BenchChatCompletionRequest(ChatCompletionRequest):
     pass
 
 
+class AddCustomModelParams(BaseModel):
+    model_id: ModelId
+
+
+class HuggingFaceSearchResult(BaseModel):
+    id: str
+    author: str = ""
+    downloads: int = 0
+    likes: int = 0
+    last_modified: str = ""
+    tags: list[str] = Field(default_factory=list)
+
+
 class PlaceInstanceParams(BaseModel):
     model_id: ModelId
     sharding: Sharding = Sharding.Pipeline
@@ -283,28 +299,7 @@ class BenchImageGenerationTaskParams(ImageGenerationTaskParams):
 
 
 class ImageEditsTaskParams(BaseModel):
-    image: UploadFile
-    prompt: str
-    background: str | None = None
-    input_fidelity: float | None = None
-    mask: UploadFile | None = None
-    model: str
-    n: int | None = 1
-    output_compression: int | None = None
-    output_format: Literal["png", "jpeg", "webp"] = "png"
-    partial_images: int | None = 0
-    quality: Literal["high", "medium", "low"] | None = "medium"
-    response_format: Literal["url", "b64_json"] | None = "b64_json"
-    size: str | None = "1024x1024"
-    stream: bool | None = False
-    user: str | None = None
-    advanced_params: AdvancedImageParams | None = None
-    # Internal flag for benchmark mode - set by API, preserved through serialization
-    bench: bool = False
-
-
-class ImageEditsInternalParams(BaseModel):
-    """Serializable version of ImageEditsTaskParams for distributed task execution."""
+    """Internal task params for image-editing requests."""
 
     image_data: str = ""  # Base64-encoded image (empty when using chunked transfer)
     total_input_chunks: int = 0
@@ -373,3 +368,45 @@ class StartDownloadResponse(CamelCaseModel):
 
 class DeleteDownloadResponse(CamelCaseModel):
     command_id: CommandId
+
+
+class TraceEventResponse(CamelCaseModel):
+    name: str
+    start_us: int
+    duration_us: int
+    rank: int
+    category: str
+
+
+class TraceResponse(CamelCaseModel):
+    task_id: str
+    traces: list[TraceEventResponse]
+
+
+class TraceCategoryStats(CamelCaseModel):
+    total_us: int
+    count: int
+    min_us: int
+    max_us: int
+    avg_us: float
+
+
+class TraceRankStats(CamelCaseModel):
+    by_category: dict[str, TraceCategoryStats]
+
+
+class TraceStatsResponse(CamelCaseModel):
+    task_id: str
+    total_wall_time_us: int
+    by_category: dict[str, TraceCategoryStats]
+    by_rank: dict[int, TraceRankStats]
+
+
+class TraceListItem(CamelCaseModel):
+    task_id: str
+    created_at: str
+    file_size: int
+
+
+class TraceListResponse(CamelCaseModel):
+    traces: list[TraceListItem]

src/exo/shared/types/claude_api.py

@@ -0,0 +1,214 @@
+"""Claude Messages API types for request/response conversion."""
+
+from typing import Any, Literal
+
+from pydantic import BaseModel, Field
+
+from exo.shared.types.common import ModelId
+
+# Tool definition types
+ClaudeToolInputSchema = dict[str, Any]
+
+
+class ClaudeToolDefinition(BaseModel, frozen=True):
+    """Tool definition in Claude Messages API request."""
+
+    name: str
+    description: str | None = None
+    input_schema: ClaudeToolInputSchema
+
+
+# Type aliases
+ClaudeRole = Literal["user", "assistant"]
+ClaudeStopReason = Literal["end_turn", "max_tokens", "stop_sequence", "tool_use"]
+
+
+# Content block types
+class ClaudeTextBlock(BaseModel, frozen=True):
+    """Text content block in Claude Messages API."""
+
+    type: Literal["text"] = "text"
+    text: str
+
+
+class ClaudeImageSource(BaseModel, frozen=True):
+    """Image source for Claude image blocks."""
+
+    type: Literal["base64", "url"]
+    media_type: str | None = None
+    data: str | None = None
+    url: str | None = None
+
+
+class ClaudeImageBlock(BaseModel, frozen=True):
+    """Image content block in Claude Messages API."""
+
+    type: Literal["image"] = "image"
+    source: ClaudeImageSource
+
+
+class ClaudeToolUseBlock(BaseModel, frozen=True):
+    """Tool use content block in Claude Messages API."""
+
+    type: Literal["tool_use"] = "tool_use"
+    id: str
+    name: str
+    input: dict[str, Any]
+
+
+class ClaudeToolResultBlock(BaseModel, frozen=True):
+    """Tool result content block in Claude Messages API request."""
+
+    type: Literal["tool_result"] = "tool_result"
+    tool_use_id: str
+    content: str | list[ClaudeTextBlock] | None = None
+    is_error: bool | None = None
+    cache_control: dict[str, str] | None = None
+
+
+ClaudeContentBlock = ClaudeTextBlock | ClaudeImageBlock | ClaudeToolUseBlock
+
+# Input content blocks can also include tool_result (sent by user after tool_use)
+ClaudeInputContentBlock = (
+    ClaudeTextBlock | ClaudeImageBlock | ClaudeToolUseBlock | ClaudeToolResultBlock
+)
+
+
+# Request types
+class ClaudeMessage(BaseModel, frozen=True):
+    """Message in Claude Messages API request."""
+
+    role: ClaudeRole
+    content: str | list[ClaudeInputContentBlock]
+
+
+class ClaudeMessagesRequest(BaseModel):
+    """Request body for Claude Messages API."""
+
+    model: ModelId
+    max_tokens: int
+    messages: list[ClaudeMessage]
+    system: str | list[ClaudeTextBlock] | None = None
+    stop_sequences: list[str] | None = None
+    stream: bool = False
+    temperature: float | None = None
+    top_p: float | None = None
+    top_k: int | None = None
+    tools: list[ClaudeToolDefinition] | None = None
+    metadata: dict[str, str] | None = None
+
+
+# Response types
+class ClaudeUsage(BaseModel, frozen=True):
+    """Token usage in Claude Messages API response."""
+
+    input_tokens: int
+    output_tokens: int
+
+
+class ClaudeMessagesResponse(BaseModel, frozen=True):
+    """Response body for Claude Messages API."""
+
+    id: str
+    type: Literal["message"] = "message"
+    role: Literal["assistant"] = "assistant"
+    content: list[ClaudeContentBlock]
+    model: str
+    stop_reason: ClaudeStopReason | None = None
+    stop_sequence: str | None = None
+    usage: ClaudeUsage
+
+
+# Streaming event types
+class ClaudeMessageStart(BaseModel, frozen=True):
+    """Partial message in message_start event."""
+
+    id: str
+    type: Literal["message"] = "message"
+    role: Literal["assistant"] = "assistant"
+    content: list[ClaudeTextBlock] = Field(default_factory=list)
+    model: str
+    stop_reason: ClaudeStopReason | None = None
+    stop_sequence: str | None = None
+    usage: ClaudeUsage
+
+
+class ClaudeMessageStartEvent(BaseModel, frozen=True):
+    """Event sent at start of message stream."""
+
+    type: Literal["message_start"] = "message_start"
+    message: ClaudeMessageStart
+
+
+class ClaudeContentBlockStartEvent(BaseModel, frozen=True):
+    """Event sent at start of a content block."""
+
+    type: Literal["content_block_start"] = "content_block_start"
+    index: int
+    content_block: ClaudeTextBlock | ClaudeToolUseBlock
+
+
+class ClaudeTextDelta(BaseModel, frozen=True):
+    """Delta for text content block."""
+
+    type: Literal["text_delta"] = "text_delta"
+    text: str
+
+
+class ClaudeInputJsonDelta(BaseModel, frozen=True):
+    """Delta for tool use input JSON content block."""
+
+    type: Literal["input_json_delta"] = "input_json_delta"
+    partial_json: str
+
+
+class ClaudeContentBlockDeltaEvent(BaseModel, frozen=True):
+    """Event sent for content block delta."""
+
+    type: Literal["content_block_delta"] = "content_block_delta"
+    index: int
+    delta: ClaudeTextDelta | ClaudeInputJsonDelta
+
+
+class ClaudeContentBlockStopEvent(BaseModel, frozen=True):
+    """Event sent at end of a content block."""
+
+    type: Literal["content_block_stop"] = "content_block_stop"
+    index: int
+
+
+class ClaudeMessageDeltaUsage(BaseModel, frozen=True):
+    """Usage in message_delta event."""
+
+    output_tokens: int
+
+
+class ClaudeMessageDelta(BaseModel, frozen=True):
+    """Delta in message_delta event."""
+
+    stop_reason: ClaudeStopReason | None = None
+    stop_sequence: str | None = None
+
+
+class ClaudeMessageDeltaEvent(BaseModel, frozen=True):
+    """Event sent with final message delta."""
+
+    type: Literal["message_delta"] = "message_delta"
+    delta: ClaudeMessageDelta
+    usage: ClaudeMessageDeltaUsage
+
+
+class ClaudeMessageStopEvent(BaseModel, frozen=True):
+    """Event sent at end of message stream."""
+
+    type: Literal["message_stop"] = "message_stop"
+
+
+ClaudeStreamEvent = (
+    ClaudeMessageStartEvent
+    | ClaudeContentBlockStartEvent
+    | ClaudeContentBlockDeltaEvent
+    | ClaudeContentBlockStopEvent
+    | ClaudeMessageDeltaEvent
+    | ClaudeMessageStopEvent
+)

src/exo/shared/types/commands.py

@@ -2,13 +2,12 @@ from pydantic import Field
 
 from exo.shared.models.model_cards import ModelCard, ModelId
 from exo.shared.types.api import (
-    BenchChatCompletionTaskParams,
-    ChatCompletionTaskParams,
-    ImageEditsInternalParams,
+    ImageEditsTaskParams,
     ImageGenerationTaskParams,
 )
 from exo.shared.types.chunks import InputImageChunk
 from exo.shared.types.common import CommandId, NodeId
+from exo.shared.types.text_generation import TextGenerationTaskParams
 from exo.shared.types.worker.instances import Instance, InstanceId, InstanceMeta
 from exo.shared.types.worker.shards import Sharding, ShardMetadata
 from exo.utils.pydantic_ext import CamelCaseModel, TaggedModel
@@ -22,16 +21,16 @@ class TestCommand(BaseCommand):
     __test__ = False
 
 
-class ChatCompletion(BaseCommand):
-    request_params: ChatCompletionTaskParams | BenchChatCompletionTaskParams
+class TextGeneration(BaseCommand):
+    task_params: TextGenerationTaskParams
 
 
 class ImageGeneration(BaseCommand):
-    request_params: ImageGenerationTaskParams
+    task_params: ImageGenerationTaskParams
 
 
 class ImageEdits(BaseCommand):
-    request_params: ImageEditsInternalParams
+    task_params: ImageEditsTaskParams
 
 
 class PlaceInstance(BaseCommand):
@@ -79,7 +78,7 @@ DownloadCommand = StartDownload | DeleteDownload
 Command = (
     TestCommand
     | RequestEventLog
-    | ChatCompletion
+    | TextGeneration
     | ImageGeneration
     | ImageEdits
     | PlaceInstance

src/exo/shared/types/events.py

@@ -1,4 +1,5 @@
 from datetime import datetime
+from typing import final
 
 from pydantic import Field
 
@@ -10,7 +11,7 @@ from exo.shared.types.worker.downloads import DownloadProgress
 from exo.shared.types.worker.instances import Instance, InstanceId
 from exo.shared.types.worker.runners import RunnerId, RunnerStatus
 from exo.utils.info_gatherer.info_gatherer import GatheredInfo
-from exo.utils.pydantic_ext import CamelCaseModel, TaggedModel
+from exo.utils.pydantic_ext import CamelCaseModel, FrozenModel, TaggedModel
 
 
 class EventId(Id):
@@ -109,6 +110,28 @@ class TopologyEdgeDeleted(BaseEvent):
     conn: Connection
 
 
+@final
+class TraceEventData(FrozenModel):
+    name: str
+    start_us: int
+    duration_us: int
+    rank: int
+    category: str
+
+
+@final
+class TracesCollected(BaseEvent):
+    task_id: TaskId
+    rank: int
+    traces: list[TraceEventData]
+
+
+@final
+class TracesMerged(BaseEvent):
+    task_id: TaskId
+    traces: list[TraceEventData]
+
+
 Event = (
     TestEvent
     | TaskCreated
@@ -127,6 +150,8 @@ Event = (
     | InputChunkReceived
     | TopologyEdgeCreated
     | TopologyEdgeDeleted
+    | TracesCollected
+    | TracesMerged
 )