fix(vllm-cpp): pass Darwin warning flag to Objective-C

The Metal source that triggers Apple Clang's constant-folding diagnostic is compiled as Objective-C. Pass the targeted suppression through CMAKE_OBJC_FLAGS as well as the C++ language flags so vllm.cpp's -Werror no longer promotes it.

Assisted-by: Codex:gpt-5

.agents/api-endpoints-and-auth.md

@@ -304,7 +304,9 @@ React pages that want to filter the ModelSelector by capability import this symb
 
 ### 4. `docs/content/` (user-facing documentation)
 
-A new capability deserves its own page under `docs/content/features/`, plus cross-links from related features and an entry in `docs/content/whats-new.md`. See the pattern used by `face-recognition.md` / `object-detection.md`.
+A new capability deserves its own page under `docs/content/features/`, plus cross-links from related features. See the pattern used by `face-recognition.md` / `object-detection.md`.
+
+Announcing it is the release's job, not this page's: the capability gets covered in the release blog post under `website/content/blog/`. See [preparing-a-release.md](preparing-a-release.md). `docs/content/whats-new.md` is only a pointer at the blog and GitHub Releases, so there is nothing to add there.
 
 ## Path protection rules
 
@@ -334,7 +336,7 @@ When adding a new endpoint:
 - [ ] Swagger block on the handler: `@Summary`, `@Tags`, `@Param`, `@Success`, `@Router`
 - [ ] If new capability area (new swagger tag): entry in `instructionDefs` in `core/http/endpoints/localai/api_instructions.go` + test count bumped in `api_instructions_test.go`
 - [ ] If new `FLAG_*` usecase flag: matching `CAP_*` symbol exported from `core/http/react-ui/src/utils/capabilities.js`
-- [ ] `docs/content/features/<feature>.md` created; cross-links from related feature pages; entry in `docs/content/whats-new.md`
+- [ ] `docs/content/features/<feature>.md` created; cross-links from related feature pages; capability covered in the release blog post (see [preparing-a-release.md](preparing-a-release.md))
 
 **Quality**
 - [ ] Error responses use `schema.ErrorResponse` format (or `echo.NewHTTPError` with a mapped gRPC status — see the `mapBackendError` helper in `core/http/endpoints/localai/images.go`)

README.md

@@ -195,7 +195,7 @@ For more details, see the [Getting Started guide](https://localai.io/basics/gett
 - **August 2025**: MLX, MLX-VLM, Diffusers, llama.cpp now supported on Apple Silicon
 - **July 2025**: All backends migrated outside the main binary — [lightweight, modular architecture](https://github.com/mudler/LocalAI/releases/tag/v3.2.0)
 
-For older news and full release notes, see [GitHub Releases](https://github.com/mudler/LocalAI/releases) and the [News page](https://localai.io/basics/news/).
+For older news and full release notes, see [GitHub Releases](https://github.com/mudler/LocalAI/releases) and the [blog](https://localai.io/blog/).
 
 ## Features
 
@@ -260,7 +260,7 @@ We also maintain [apex-quant](https://github.com/localai-org/apex-quant), a per-
 - [Kubernetes installation](https://localai.io/basics/getting_started/#run-localai-in-kubernetes)
 - [Integrations & community projects](https://localai.io/docs/integrations/)
 - [Installation video walkthrough](https://www.youtube.com/watch?v=cMVNnlqwfw4)
-- [Media & blog posts](https://localai.io/basics/news/#media-blogs-social)
+- [Blog: release write-ups, benchmarks and engineering notes](https://localai.io/blog/)
 - [Examples](https://github.com/mudler/LocalAI-examples) — including the [realtime voice assistant demo](https://github.com/localai-org/localai-realtime-demo) (Go client for the Realtime API with tool calling)
 
 ## Team

backend/cpp/audio-cpp/Makefile

@@ -9,7 +9,7 @@
 # recipe is a make target (not a prepare.sh) so 'make purge && make' is a clean
 # rebuild and so the bump bot can see the pin.
 
-AUDIO_CPP_VERSION?=5a8312ef7b8aa7cf14e9a24ac568cabd8725d68a
+AUDIO_CPP_VERSION?=4e3aea2fd99aeaa5924e71c51eb2793846045332
 AUDIO_CPP_REPO?=https://github.com/0xShug0/audio.cpp
 
 CURRENT_MAKEFILE_DIR := $(dir $(abspath $(lastword $(MAKEFILE_LIST))))

backend/cpp/ik-llama-cpp/Makefile

@@ -1,5 +1,5 @@
 
-IK_LLAMA_VERSION?=cb9147fd0d9c08a9a84eee5ac405a73f4e10e3e1
+IK_LLAMA_VERSION?=60389410a1ff01f9d37dcc6261db33b3183bdea2
 LLAMA_REPO?=https://github.com/ikawrakow/ik_llama.cpp
 
 CMAKE_ARGS?=

backend/cpp/llama-cpp/Makefile

@@ -1,5 +1,5 @@
 
-LLAMA_VERSION?=a7a6d0d269c896218b6c78e0933bd6a17519d3f6
+LLAMA_VERSION?=221f0f6356efe2260023208365705ec5d5a7c8f5
 LLAMA_REPO?=https://github.com/ggerganov/llama.cpp
 
 CMAKE_ARGS?=

backend/go/crispasr/Makefile

@@ -8,7 +8,7 @@ JOBS?=$(shell nproc --ignore=1)
 
 # CrispASR version (release tag)
 CRISPASR_REPO?=https://github.com/CrispStrobe/CrispASR
-CRISPASR_VERSION?=fcb79282a6bc52e13d858026c42b24fb6e63c97a
+CRISPASR_VERSION?=fe3caf8e363b27572dbdd1a9d37083f25e6decda
 SO_TARGET?=libgocrispasr.so
 
 CMAKE_ARGS+=-DBUILD_SHARED_LIBS=OFF

backend/go/vllm-cpp/Makefile

@@ -11,7 +11,7 @@ JOBS?=$(shell nproc --ignore=1 2>/dev/null || sysctl -n hw.ncpu 2>/dev/null || e
 
 # vllm.cpp version
 VLLM_CPP_REPO?=https://github.com/mudler/vllm.cpp
-VLLM_CPP_VERSION?=9e1c9025ae61167a3335454d7cc0de6093c21845
+VLLM_CPP_VERSION?=a42b8187caff02c570c28e19e4dc2b1d7f55ed14
 
 # The backend consumes only the stable C ABI (libvllm + include/vllm.h), so the
 # server, examples and tests of the engine are never built here.
@@ -56,6 +56,12 @@ endif
 UNAME_S := $(shell uname -s)
 ifeq ($(UNAME_S),Darwin)
 	LIB=libvllm.dylib
+	# Apple Clang diagnoses a pair of constant-folded array bounds in the Metal
+	# build as a GNU extension. Disable that diagnostic for both Objective-C and
+	# C++ because vllm.cpp appends target-local -Werror after these global flags.
+	CMAKE_ARGS+=-DCMAKE_CXX_FLAGS=-Wno-gnu-folding-constant
+	CMAKE_ARGS+=-DCMAKE_OBJC_FLAGS=-Wno-gnu-folding-constant
+	CMAKE_ARGS+=-DCMAKE_OBJCXX_FLAGS=-Wno-gnu-folding-constant
 else
 	LIB=libvllm.so
 endif

backend/go/vllm-cpp/backend.go

@@ -109,6 +109,16 @@ func (v *VllmCpp) Load(opts *pb.ModelOptions) error {
 
 	v.opts = parseOptions(opts)
 
+	// A DFlash draft is a second checkpoint the engine opens by path, and the
+	// engine never downloads one. Resolve it against LocalAI's models directory
+	// now so a repo-id spelling works, and so a missing draft fails here with an
+	// actionable message rather than as an HF-cache miss inside the load.
+	resolvedSpec, err := resolveDraftModelPath(v.opts.speculativeConfig, opts.ModelPath)
+	if err != nil {
+		return err
+	}
+	v.opts.speculativeConfig = resolvedSpec
+
 	mp := defaultModelParams()
 	if v.opts.blockSize > 0 {
 		mp.BlockSize = v.opts.blockSize
@@ -116,34 +126,62 @@ func (v *VllmCpp) Load(opts *pb.ModelOptions) error {
 	if v.opts.numBlocks > 0 {
 		mp.NumBlocks = v.opts.numBlocks
 	}
+	// Sequence-length precedence, narrowest source last: context_size is the
+	// generic LocalAI knob every backend honours, max_model_len is the
+	// vLLM-specific one, and engine_args.max_model_len is the explicit
+	// vllm-cpp override.
 	if opts.ContextSize > 0 {
 		mp.MaxModelLen = opts.ContextSize
 	}
+	if opts.MaxModelLen > 0 {
+		mp.MaxModelLen = opts.MaxModelLen
+	}
+	if v.opts.maxModelLen > 0 {
+		mp.MaxModelLen = v.opts.maxModelLen
+	}
 	if v.opts.maxNumSeqs > 0 {
 		mp.MaxNumSeqs = v.opts.maxNumSeqs
 	}
+	if v.opts.maxNumBatchedTokens > 0 {
+		mp.MaxNumBatchedTokens = v.opts.maxNumBatchedTokens
+	}
+	mp.EnablePrefixCaching = v.opts.enablePrefixCaching
+	mp.EnableJumpForward = v.opts.enableJumpForward
 
+	// Every string below is borrowed by C for the duration of the load call
+	// only (the library copies what it keeps), so the backing slices just have
+	// to outlive vllmEngineLoad - hence the single KeepAlive after it.
 	modelC := cString(model)
 	mp.ModelPath = uintptr(unsafe.Pointer(&modelC[0])) // #nosec G103 -- borrowed by C for the load call only
-	var toolParserC, reasoningParserC []byte
-	if v.opts.toolParser != "" {
-		toolParserC = cString(v.opts.toolParser)
-		mp.ToolParser = uintptr(unsafe.Pointer(&toolParserC[0])) // #nosec G103 -- borrowed by C for the load call only
-	}
-	if v.opts.reasoningParser != "" {
-		reasoningParserC = cString(v.opts.reasoningParser)
-		mp.ReasoningParser = uintptr(unsafe.Pointer(&reasoningParserC[0])) // #nosec G103 -- borrowed by C for the load call only
+	keep := [][]byte{modelC}
+	setStr := func(dst *uintptr, s string) {
+		if s == "" {
+			return
+		}
+		b := cString(s)
+		keep = append(keep, b)
+		*dst = uintptr(unsafe.Pointer(&b[0])) // #nosec G103 -- borrowed by C for the load call only
 	}
+	setStr(&mp.ToolParser, v.opts.toolParser)
+	setStr(&mp.ReasoningParser, v.opts.reasoningParser)
+	setStr(&mp.SpeculativeConfig, v.opts.speculativeConfig)
+	setStr(&mp.KVTransferConfig, v.opts.kvTransferConfig)
+	setStr(&mp.SchedulingPolicy, v.opts.schedulingPolicy)
+	setStr(&mp.TokenizerConfigPath, v.opts.tokenizerConfigPath)
 
 	xlog.Info("[vllm-cpp] Load", "model", model, "engine", vllmVersion(),
 		"blockSize", mp.BlockSize, "numBlocks", mp.NumBlocks,
-		"maxModelLen", mp.MaxModelLen, "maxNumSeqs", mp.MaxNumSeqs)
+		"maxModelLen", mp.MaxModelLen, "maxNumSeqs", mp.MaxNumSeqs,
+		"maxNumBatchedTokens", mp.MaxNumBatchedTokens,
+		"prefixCaching", triStateName(mp.EnablePrefixCaching),
+		"jumpForward", triStateName(mp.EnableJumpForward),
+		"schedulingPolicy", v.opts.schedulingPolicy,
+		"speculativeConfig", v.opts.speculativeConfig,
+		"kvTransferConfig", v.opts.kvTransferConfig)
 
 	var engine uintptr
 	rc := vllmEngineLoad(unsafe.Pointer(&mp), unsafe.Pointer(&engine)) // #nosec G103 -- POD out-params
-	runtime.KeepAlive(modelC)
-	runtime.KeepAlive(toolParserC)
-	runtime.KeepAlive(reasoningParserC)
+	runtime.KeepAlive(keep)
 	if rc != vllmOK {
 		return fmt.Errorf("vllm-cpp: engine load failed: %s", vllmLastError())
 	}

backend/go/vllm-cpp/govllmcpp.go

@@ -1,6 +1,6 @@
 package main
 
-// purego bindings for the vllm.cpp stable C ABI (include/vllm.h, ABI v2).
+// purego bindings for the vllm.cpp stable C ABI (include/vllm.h, ABI v10).
 //
 // The structs below are hand-mirrored PODs of the C declarations, with
 // explicit padding so the Go layout matches the C layout on linux/darwin
@@ -18,23 +18,56 @@ import (
 )
 
 // abiVersion is the VLLM_ABI_VERSION this file mirrors (vllm.h).
-const abiVersion = 5
+const abiVersion = 10
+
+// The ABI's tri-state toggles (enable_prefix_caching ABI v7,
+// enable_jump_forward ABI v10) share one encoding: 0 is NOT "off", it is
+// "defer" - to the model capability for prefix caching, to the environment for
+// jump forward. Only 2 is an explicit off.
+const (
+	triStateDefer int32 = 0
+	triStateOn    int32 = 1
+	triStateOff   int32 = 2
+)
+
+// triStateName renders a tri-state for the load log line, where "0" would
+// otherwise read as "off" rather than "whatever the default resolves to".
+func triStateName(state int32) string {
+	switch state {
+	case triStateOn:
+		return "on"
+	case triStateOff:
+		return "off"
+	default:
+		return "model-default"
+	}
+}
 
 // vllm_status (vllm.h).
 const (
 	vllmOK = 0
 )
 
-// cModelParams mirrors vllm_model_params.
+// cModelParams mirrors vllm_model_params. The int32 fields sit in pairs so the
+// interior needs no padding on LP64, but the struct is 8-aligned (it holds
+// pointers) and ends on a lone int32, so the trailing pad is explicit. Offsets
+// and total size are asserted in vllmcpp_test.go.
 type cModelParams struct {
 	ModelPath           uintptr // const char*
-	TokenizerConfigPath uintptr // const char*
+	TokenizerConfigPath uintptr // const char*; NULL = <model_dir>/... (ABI v9)
 	BlockSize           int32
 	NumBlocks           int32
 	MaxModelLen         int32
 	MaxNumSeqs          int32
 	ToolParser          uintptr // const char*; NULL = auto-detect (ABI v4)
 	ReasoningParser     uintptr // const char*; NULL = auto-detect (ABI v5)
+	SpeculativeConfig   uintptr // const char* JSON; NULL = no speculation (ABI v6)
+	EnablePrefixCaching int32   // tri-state 0/1/2 (ABI v7)
+	MaxNumBatchedTokens int32   // <= 0 = per-arch default (ABI v9)
+	SchedulingPolicy    uintptr // const char*; NULL = "fcfs" (ABI v9)
+	KVTransferConfig    uintptr // const char* JSON; NULL = no connector (ABI v9)
+	EnableJumpForward   int32   // tri-state 0/1/2 (ABI v10)
+	_                   [4]byte // trailing pad to the struct's 8-byte alignment
 }
 
 // cSamplingParams mirrors vllm_sampling_params (ABI v2, structured fields
@@ -65,6 +98,12 @@ type cSamplingParams struct {
 	StructuredGrammar    uintptr // const char*
 	StructuredJSONObject int32
 	_                    [4]byte
+	// ABI v8 tail. LocalAI installs no custom logits processor, but the fields
+	// MUST be mirrored: the C side reads them off the pointer we hand it, so a
+	// Go struct that stopped at StructuredJSONObject would have the engine read
+	// 16 bytes past our allocation and call whatever garbage sat there.
+	LogitsProcessor         uintptr // vllm_logits_processor; NULL = none
+	LogitsProcessorUserData uintptr // void*
 }
 
 // cCompletion mirrors vllm_completion.

backend/go/vllm-cpp/options.go

@@ -1,30 +1,80 @@
 package main
 
-// Engine-sizing knobs carried through the model config's free-form
-// `options:` list ("key:value" entries), mirroring how the other in-house
-// backends pass engine-specific settings that have no proto field.
+// Load-time engine configuration, from two config surfaces:
+//
+//   - `engine_args:` (ModelOptions.EngineArgs, a JSON object) is the canonical
+//     one. Keys are spelled exactly as vLLM's own CLI flags, so a config written
+//     against vLLM works verbatim here - `speculative_config` and
+//     `kv_transfer_config` in particular take the same JSON documents vLLM's
+//     --speculative-config / --kv-transfer-config accept, and are handed to the
+//     engine unparsed.
+//   - `options:` (the free-form "key:value" list) is the older surface this
+//     backend shipped with. It is still honoured so existing configs keep
+//     working; engine_args wins on any key set in both.
+//
+// Anything unrecognised is ignored rather than fatal: the engine validates the
+// documents it is given and reports a precise error at load, and a config that
+// also carries knobs for a different backend must not fail the load here.
 
 import (
+	"encoding/json"
+	"fmt"
+	"os"
+	"path"
+	"path/filepath"
 	"strconv"
 	"strings"
 
 	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
+	"github.com/mudler/xlog"
 )
 
 type loadOptions struct {
 	blockSize  int32 // KV block size (tokens/block); engine default 32.
 	numBlocks  int32 // KV blocks to allocate; engine default 256.
 	maxNumSeqs int32 // max concurrent sequences; engine default 8.
+	// Max sequence length. Also settable through the model config's
+	// context_size / max_model_len; see Load for the precedence.
+	maxModelLen int32
+	// Per-step chunked-prefill token budget (ABI v9). 0 = the engine's
+	// bounded per-arch default.
+	maxNumBatchedTokens int32
+	// Automatic prefix caching tri-state (ABI v7): 0 = the model-capability
+	// default, 1 = force on, 2 = force off.
+	enablePrefixCaching int32
+	// Jump-forward decoding tri-state (ABI v10), SGLang's grammar-speed subset:
+	// 0 = defer to the environment (VT_ENABLE_JUMP_FORWARD, default off),
+	// 1 = force on, 2 = force off.
+	enableJumpForward int32
+	// Scheduler admission policy (ABI v9): "" = fcfs, else fcfs|priority|lpm.
+	schedulingPolicy string
 	// Engine-side parser selection (ABI v4/v5). Empty = the engine
 	// auto-detects from the chat template; "none" disables the reasoning
 	// split; unknown names fail the first chat call.
 	toolParser      string
 	reasoningParser string
+	// Speculative decoding (ABI v6), as vLLM's --speculative-config JSON:
+	// {"method":"mtp"|"dflash"|"ngram", ...}. Empty = no speculation.
+	speculativeConfig string
+	// External KV connector / LMCache (ABI v9), as vLLM's --kv-transfer-config
+	// JSON. Empty = no connector.
+	kvTransferConfig string
+	// Override for the tokenizer_config.json the chat template is read from
+	// (ABI v9). Empty = <model_dir>/tokenizer_config.json.
+	tokenizerConfigPath string
 }
 
 func parseOptions(opts *pb.ModelOptions) loadOptions {
 	lo := loadOptions{}
-	for _, o := range opts.GetOptions() {
+	applyOptionsList(&lo, opts.GetOptions())
+	applyEngineArgs(&lo, opts.GetEngineArgs())
+	return lo
+}
+
+// applyOptionsList reads the legacy free-form "key:value" list. strings.Cut
+// splits on the FIRST colon only, so a JSON object value survives intact.
+func applyOptionsList(lo *loadOptions, options []string) {
+	for _, o := range options {
 		k, v, found := strings.Cut(o, ":")
 		if !found {
 			continue
@@ -36,13 +86,211 @@ func parseOptions(opts *pb.ModelOptions) loadOptions {
 			lo.numBlocks = parseInt32(v, lo.numBlocks)
 		case "max_num_seqs":
 			lo.maxNumSeqs = parseInt32(v, lo.maxNumSeqs)
-		case "tool_parser":
+		case "max_num_batched_tokens":
+			lo.maxNumBatchedTokens = parseInt32(v, lo.maxNumBatchedTokens)
+		case "max_model_len":
+			lo.maxModelLen = parseInt32(v, lo.maxModelLen)
+		case "scheduling_policy", "schedule_policy":
+			lo.schedulingPolicy = strings.TrimSpace(v)
+		case "tool_parser", "tool_call_parser":
 			lo.toolParser = strings.TrimSpace(v)
 		case "reasoning_parser":
 			lo.reasoningParser = strings.TrimSpace(v)
+		case "speculative_config":
+			lo.speculativeConfig = strings.TrimSpace(v)
+		case "kv_transfer_config":
+			lo.kvTransferConfig = strings.TrimSpace(v)
+		case "tokenizer_config", "tokenizer_config_path":
+			lo.tokenizerConfigPath = strings.TrimSpace(v)
+		case "enable_prefix_caching", "enable_radix_attention":
+			if b, err := strconv.ParseBool(strings.TrimSpace(v)); err == nil {
+				lo.enablePrefixCaching = boolTriState(b)
+			}
+		case "enable_jump_forward":
+			if b, err := strconv.ParseBool(strings.TrimSpace(v)); err == nil {
+				lo.enableJumpForward = boolTriState(b)
+			}
 		}
 	}
-	return lo
+}
+
+// applyEngineArgs overlays the `engine_args:` JSON object. A document that does
+// not parse is logged and skipped: engine_args is shared with the other engines
+// (the vLLM and SGLang backends read the same field), so a stray key must not
+// take the model down.
+func applyEngineArgs(lo *loadOptions, engineArgs string) {
+	if strings.TrimSpace(engineArgs) == "" {
+		return
+	}
+	var args map[string]any
+	if err := json.Unmarshal([]byte(engineArgs), &args); err != nil {
+		xlog.Warn("[vllm-cpp] ignoring unparseable engine_args", "error", err)
+		return
+	}
+	for k, v := range args {
+		switch k {
+		case "block_size":
+			lo.blockSize = jsonInt32(v, lo.blockSize)
+		case "num_blocks":
+			lo.numBlocks = jsonInt32(v, lo.numBlocks)
+		case "max_num_seqs":
+			lo.maxNumSeqs = jsonInt32(v, lo.maxNumSeqs)
+		case "max_num_batched_tokens":
+			lo.maxNumBatchedTokens = jsonInt32(v, lo.maxNumBatchedTokens)
+		case "max_model_len":
+			lo.maxModelLen = jsonInt32(v, lo.maxModelLen)
+		case "scheduling_policy", "schedule_policy":
+			lo.schedulingPolicy = jsonString(v, lo.schedulingPolicy)
+		case "tool_parser", "tool_call_parser":
+			lo.toolParser = jsonString(v, lo.toolParser)
+		case "reasoning_parser":
+			lo.reasoningParser = jsonString(v, lo.reasoningParser)
+		case "tokenizer_config", "tokenizer_config_path":
+			lo.tokenizerConfigPath = jsonString(v, lo.tokenizerConfigPath)
+		case "speculative_config":
+			lo.speculativeConfig = jsonDocument(v, lo.speculativeConfig, k)
+		case "kv_transfer_config":
+			lo.kvTransferConfig = jsonDocument(v, lo.kvTransferConfig, k)
+		case "enable_prefix_caching", "enable_radix_attention":
+			if b, ok := v.(bool); ok {
+				lo.enablePrefixCaching = boolTriState(b)
+			}
+		case "enable_jump_forward":
+			if b, ok := v.(bool); ok {
+				lo.enableJumpForward = boolTriState(b)
+			}
+		default:
+			xlog.Debug("[vllm-cpp] ignoring unknown engine_args key", "key", k)
+		}
+	}
+}
+
+// boolTriState maps a YAML/JSON boolean onto the ABI's tri-state encoding. An
+// explicit `false` must reach the engine as force-OFF (2), NOT as the 0 that
+// means "defer". The difference is real in both directions: prefix caching
+// defaults ON for dense archs and OFF for hybrid ones, and jump forward defers
+// to VT_ENABLE_JUMP_FORWARD.
+func boolTriState(on bool) int32 {
+	if on {
+		return triStateOn
+	}
+	return triStateOff
+}
+
+// jsonDocument normalises an object-valued engine_args entry to a JSON string
+// for the C ABI. YAML nesting arrives as a map (the natural spelling); a
+// pre-encoded JSON string is accepted too, since a config round-tripped through
+// a flat store may carry it that way.
+func jsonDocument(v any, fallback string, key string) string {
+	switch t := v.(type) {
+	case string:
+		if strings.TrimSpace(t) == "" {
+			return fallback
+		}
+		return t
+	default:
+		buf, err := json.Marshal(t)
+		if err != nil {
+			xlog.Warn("[vllm-cpp] ignoring unencodable engine_args value", "key", key, "error", err)
+			return fallback
+		}
+		return string(buf)
+	}
+}
+
+func jsonString(v any, fallback string) string {
+	s, ok := v.(string)
+	if !ok {
+		return fallback
+	}
+	return strings.TrimSpace(s)
+}
+
+// jsonInt32 accepts the float64 a JSON number decodes to, plus the string
+// spelling a YAML config may produce. Non-positive values keep the fallback:
+// every knob this covers uses "<= 0 means the engine default".
+func jsonInt32(v any, fallback int32) int32 {
+	switch t := v.(type) {
+	case float64:
+		if t <= 0 || t > 1<<31-1 {
+			return fallback
+		}
+		return int32(t)
+	case string:
+		return parseInt32(t, fallback)
+	default:
+		return fallback
+	}
+}
+
+// resolveDraftModelPath rewrites a DFlash draft reference into an absolute path
+// the engine can actually open.
+//
+// The engine resolves `speculative_config.model` against a directory containing
+// config.json, or against ~/.cache/huggingface/hub/models--<org>--<repo>/
+// snapshots/* - and it NEVER downloads. LocalAI keeps models in its own
+// directory, so a bare HF repo id (the spelling the vLLM docs teach) misses the
+// HF cache and dies deep in the load with "draft checkpoint not found", which
+// reads like a broken checkpoint rather than a missing download.
+//
+// So: try the reference as given, then the last path segment under the models
+// dir (`z-lab/Qwen3.6-27B-DFlash` -> `<models>/Qwen3.6-27B-DFlash`, which is
+// what LocalAI's own downloader produces), then the whole reference under the
+// models dir. If none exist, fail HERE with a message naming both what was
+// asked for and where we looked.
+//
+// mtp and ngram carry no separate draft checkpoint, so they pass through. A
+// document that does not parse also passes through: the engine owns config
+// validation and produces the better error.
+func resolveDraftModelPath(speculativeConfig, modelsDir string) (string, error) {
+	if strings.TrimSpace(speculativeConfig) == "" {
+		return speculativeConfig, nil
+	}
+	var spec map[string]any
+	if err := json.Unmarshal([]byte(speculativeConfig), &spec); err != nil {
+		return speculativeConfig, nil
+	}
+	if method, _ := spec["method"].(string); !strings.EqualFold(method, "dflash") {
+		return speculativeConfig, nil
+	}
+
+	ref, _ := spec["model"].(string)
+	ref = strings.TrimSpace(ref)
+	if ref == "" {
+		return "", fmt.Errorf(
+			"vllm-cpp: speculative_config method %q requires a \"model\" key naming the draft checkpoint", "dflash")
+	}
+
+	candidates := []string{ref}
+	if modelsDir != "" {
+		if base := path.Base(filepath.ToSlash(ref)); base != "" && base != "." && base != "/" {
+			candidates = append(candidates, filepath.Join(modelsDir, base))
+		}
+		candidates = append(candidates, filepath.Join(modelsDir, filepath.FromSlash(ref)))
+	}
+
+	for _, c := range candidates {
+		if _, err := os.Stat(filepath.Join(c, "config.json")); err != nil {
+			continue
+		}
+		abs, err := filepath.Abs(c)
+		if err != nil {
+			abs = c
+		}
+		spec["model"] = abs
+		out, err := json.Marshal(spec)
+		if err != nil {
+			return "", fmt.Errorf("vllm-cpp: re-encoding speculative_config: %w", err)
+		}
+		xlog.Info("[vllm-cpp] resolved DFlash draft checkpoint", "reference", ref, "path", abs)
+		return string(out), nil
+	}
+
+	return "", fmt.Errorf(
+		"vllm-cpp: DFlash draft checkpoint %q not found (looked in: %s). "+
+			"The engine does not download drafts - install the draft model into LocalAI first, "+
+			"or set speculative_config.model to an absolute path to a directory containing config.json",
+		ref, strings.Join(candidates, ", "))
 }
 
 func parseInt32(s string, fallback int32) int32 {

backend/go/vllm-cpp/vllmcpp_test.go

@@ -16,7 +16,7 @@ func TestVllmCpp(t *testing.T) {
 	RunSpecs(t, "vllm-cpp suite")
 }
 
-// The Go POD mirrors must match the C struct layout of vllm.h (ABI v2)
+// The Go POD mirrors must match the C struct layout of vllm.h (ABI v9)
 // byte-for-byte: these offsets are the C offsets on LP64 (linux/darwin
 // amd64+arm64). A failure here means govllmcpp.go drifted from vllm.h.
 var _ = Describe("C ABI struct mirrors", func() {
@@ -30,10 +30,18 @@ var _ = Describe("C ABI struct mirrors", func() {
 		Expect(unsafe.Offsetof(p.MaxNumSeqs)).To(Equal(uintptr(28)))
 		Expect(unsafe.Offsetof(p.ToolParser)).To(Equal(uintptr(32)))
 		Expect(unsafe.Offsetof(p.ReasoningParser)).To(Equal(uintptr(40)))
-		Expect(unsafe.Sizeof(p)).To(Equal(uintptr(48)))
+		Expect(unsafe.Offsetof(p.SpeculativeConfig)).To(Equal(uintptr(48)))
+		Expect(unsafe.Offsetof(p.EnablePrefixCaching)).To(Equal(uintptr(56)))
+		Expect(unsafe.Offsetof(p.MaxNumBatchedTokens)).To(Equal(uintptr(60)))
+		Expect(unsafe.Offsetof(p.SchedulingPolicy)).To(Equal(uintptr(64)))
+		Expect(unsafe.Offsetof(p.KVTransferConfig)).To(Equal(uintptr(72)))
+		Expect(unsafe.Offsetof(p.EnableJumpForward)).To(Equal(uintptr(80)))
+		// 88, not 84: the struct is 8-aligned (it holds pointers), so the
+		// trailing int32 is padded out. Go pads identically.
+		Expect(unsafe.Sizeof(p)).To(Equal(uintptr(88)))
 	})
 
-	It("cSamplingParams matches vllm_sampling_params (ABI v2)", func() {
+	It("cSamplingParams matches vllm_sampling_params (ABI v8)", func() {
 		var p cSamplingParams
 		Expect(unsafe.Offsetof(p.Temperature)).To(Equal(uintptr(0)))
 		Expect(unsafe.Offsetof(p.TopP)).To(Equal(uintptr(4)))
@@ -55,7 +63,9 @@ var _ = Describe("C ABI struct mirrors", func() {
 		Expect(unsafe.Offsetof(p.NStructuredChoice)).To(Equal(uintptr(96)))
 		Expect(unsafe.Offsetof(p.StructuredGrammar)).To(Equal(uintptr(104)))
 		Expect(unsafe.Offsetof(p.StructuredJSONObject)).To(Equal(uintptr(112)))
-		Expect(unsafe.Sizeof(p)).To(Equal(uintptr(120)))
+		Expect(unsafe.Offsetof(p.LogitsProcessor)).To(Equal(uintptr(120)))
+		Expect(unsafe.Offsetof(p.LogitsProcessorUserData)).To(Equal(uintptr(128)))
+		Expect(unsafe.Sizeof(p)).To(Equal(uintptr(136)))
 	})
 
 	It("cCompletion matches vllm_completion", func() {
@@ -68,6 +78,23 @@ var _ = Describe("C ABI struct mirrors", func() {
 	})
 })
 
+// Pin/mirror skew is the failure mode this backend is most exposed to: the Go
+// PODs above are hand-written against one VLLM_ABI_VERSION, and the Makefile
+// pins the vllm.cpp commit that produces it. This spec catches drift without
+// needing model weights - set VLLM_CPP_LIBRARY to a built libvllm and it binds
+// every symbol and compares the library's reported ABI against the mirrors'.
+var _ = Describe("real library ABI handshake", func() {
+	It("binds every symbol and reports the ABI the mirrors were written against", func() {
+		lib := os.Getenv("VLLM_CPP_LIBRARY")
+		if lib == "" {
+			Skip("VLLM_CPP_LIBRARY not set; skipping the real-library handshake")
+		}
+		Expect(registerLib(lib)).To(Succeed())
+		Expect(vllmABIVersion()).To(Equal(int32(abiVersion)))
+		Expect(vllmVersion()).NotTo(BeEmpty())
+	})
+})
+
 var _ = Describe("parseOptions", func() {
 	It("extracts the engine sizing knobs", func() {
 		lo := parseOptions(&pb.ModelOptions{Options: []string{
@@ -83,6 +110,129 @@ var _ = Describe("parseOptions", func() {
 		}})
 		Expect(lo).To(Equal(loadOptions{}))
 	})
+
+	It("carries a speculative_config JSON value through the legacy options list", func() {
+		// strings.Cut splits on the FIRST colon only, so a JSON object value
+		// survives the "key:value" spelling intact.
+		lo := parseOptions(&pb.ModelOptions{Options: []string{
+			`speculative_config:{"method":"mtp","num_speculative_tokens":1}`,
+		}})
+		Expect(lo.speculativeConfig).To(Equal(`{"method":"mtp","num_speculative_tokens":1}`))
+	})
+})
+
+var _ = Describe("engine_args", func() {
+	It("maps every load knob onto the C model params", func() {
+		lo := parseOptions(&pb.ModelOptions{EngineArgs: `{
+			"block_size": 64,
+			"num_blocks": 1024,
+			"max_model_len": 16384,
+			"max_num_seqs": 32,
+			"max_num_batched_tokens": 8192,
+			"enable_prefix_caching": true,
+			"scheduling_policy": "lpm",
+			"tool_parser": "qwen3",
+			"reasoning_parser": "deepseek_r1",
+			"tokenizer_config": "/models/tok/tokenizer_config.json"
+		}`})
+		Expect(lo.blockSize).To(Equal(int32(64)))
+		Expect(lo.numBlocks).To(Equal(int32(1024)))
+		Expect(lo.maxModelLen).To(Equal(int32(16384)))
+		Expect(lo.maxNumSeqs).To(Equal(int32(32)))
+		Expect(lo.maxNumBatchedTokens).To(Equal(int32(8192)))
+		Expect(lo.enablePrefixCaching).To(Equal(int32(1)))
+		Expect(lo.schedulingPolicy).To(Equal("lpm"))
+		Expect(lo.toolParser).To(Equal("qwen3"))
+		Expect(lo.reasoningParser).To(Equal("deepseek_r1"))
+		Expect(lo.tokenizerConfigPath).To(Equal("/models/tok/tokenizer_config.json"))
+	})
+
+	It("re-marshals a nested speculative_config object to JSON for the engine", func() {
+		lo := parseOptions(&pb.ModelOptions{EngineArgs: `{
+			"speculative_config": {"method": "mtp", "num_speculative_tokens": 1}
+		}`})
+		Expect(lo.speculativeConfig).To(MatchJSON(`{"method":"mtp","num_speculative_tokens":1}`))
+	})
+
+	It("re-marshals a nested kv_transfer_config object (LMCache) to JSON", func() {
+		lo := parseOptions(&pb.ModelOptions{EngineArgs: `{
+			"kv_transfer_config": {
+				"kv_connector": "LMCacheConnector",
+				"kv_role": "kv_both",
+				"kv_connector_extra_config": {"host": "127.0.0.1", "port": 65432}
+			}
+		}`})
+		Expect(lo.kvTransferConfig).To(MatchJSON(`{
+			"kv_connector":"LMCacheConnector",
+			"kv_role":"kv_both",
+			"kv_connector_extra_config":{"host":"127.0.0.1","port":65432}
+		}`))
+	})
+
+	It("accepts a pre-encoded JSON string for the object-valued knobs", func() {
+		// A config written by hand (or round-tripped through a flat store) may
+		// carry the object as a string; both spellings reach the engine the same.
+		lo := parseOptions(&pb.ModelOptions{EngineArgs: `{
+			"speculative_config": "{\"method\":\"ngram\",\"num_speculative_tokens\":4}"
+		}`})
+		Expect(lo.speculativeConfig).To(MatchJSON(`{"method":"ngram","num_speculative_tokens":4}`))
+	})
+
+	It("maps enable_prefix_caching false onto the force-OFF tri-state", func() {
+		// The C ABI tri-state is 0=model default, 1=on, 2=off, so an explicit
+		// `false` must NOT collapse to the 0 that means "let the model decide".
+		lo := parseOptions(&pb.ModelOptions{EngineArgs: `{"enable_prefix_caching": false}`})
+		Expect(lo.enablePrefixCaching).To(Equal(int32(2)))
+	})
+
+	It("leaves the prefix-caching tri-state at the model default when unset", func() {
+		lo := parseOptions(&pb.ModelOptions{EngineArgs: `{"max_num_seqs": 4}`})
+		Expect(lo.enablePrefixCaching).To(Equal(int32(0)))
+	})
+
+	It("accepts the radix-attention alias upstream documents for prefix caching", func() {
+		lo := parseOptions(&pb.ModelOptions{EngineArgs: `{"enable_radix_attention": true}`})
+		Expect(lo.enablePrefixCaching).To(Equal(int32(1)))
+	})
+
+	It("maps enable_jump_forward onto its own tri-state", func() {
+		// ABI v10. Same tri-state shape as prefix caching, and the same trap:
+		// an explicit false must be force-OFF (2), not the 0 that defers to the
+		// environment.
+		on := parseOptions(&pb.ModelOptions{EngineArgs: `{"enable_jump_forward": true}`})
+		Expect(on.enableJumpForward).To(Equal(int32(1)))
+		off := parseOptions(&pb.ModelOptions{EngineArgs: `{"enable_jump_forward": false}`})
+		Expect(off.enableJumpForward).To(Equal(int32(2)))
+		unset := parseOptions(&pb.ModelOptions{EngineArgs: `{"max_num_seqs": 4}`})
+		Expect(unset.enableJumpForward).To(Equal(int32(0)))
+	})
+
+	It("reads enable_jump_forward from the legacy options list too", func() {
+		lo := parseOptions(&pb.ModelOptions{Options: []string{"enable_jump_forward:true"}})
+		Expect(lo.enableJumpForward).To(Equal(int32(1)))
+	})
+
+	It("lets engine_args override the legacy options list", func() {
+		lo := parseOptions(&pb.ModelOptions{
+			Options:    []string{"max_num_seqs:8", "block_size:16"},
+			EngineArgs: `{"max_num_seqs": 64}`,
+		})
+		Expect(lo.maxNumSeqs).To(Equal(int32(64))) // engine_args wins
+		Expect(lo.blockSize).To(Equal(int32(16)))  // untouched keys survive
+	})
+
+	It("ignores malformed engine_args rather than failing the load", func() {
+		lo := parseOptions(&pb.ModelOptions{
+			Options:    []string{"max_num_seqs:8"},
+			EngineArgs: `{not json`,
+		})
+		Expect(lo.maxNumSeqs).To(Equal(int32(8)))
+	})
+
+	It("ignores unknown keys", func() {
+		lo := parseOptions(&pb.ModelOptions{EngineArgs: `{"gpu_memory_utilization": 0.9}`})
+		Expect(lo).To(Equal(loadOptions{}))
+	})
 })
 
 var _ = Describe("samplingFromPredict", func() {
@@ -135,6 +285,91 @@ var _ = Describe("samplingFromPredict", func() {
 	})
 })
 
+// The engine resolves speculative_config.model against a local directory or
+// ~/.cache/huggingface/hub ONLY - it never downloads. LocalAI keeps models in
+// its own directory, so a bare repo id would miss the HF cache and fail deep in
+// the load with a confusing "draft checkpoint not found". Resolve it here.
+var _ = Describe("resolveDraftModelPath", func() {
+	var modelsDir string
+
+	BeforeEach(func() {
+		modelsDir = GinkgoT().TempDir()
+	})
+
+	// draftDir creates a plausible draft checkpoint under models/.
+	draftDir := func(name string) string {
+		d := filepath.Join(modelsDir, name)
+		Expect(os.MkdirAll(d, 0o750)).To(Succeed())
+		Expect(os.WriteFile(filepath.Join(d, "config.json"), []byte("{}"), 0o600)).To(Succeed())
+		return d
+	}
+
+	It("rewrites a repo id to the matching directory in the models dir", func() {
+		want := draftDir("Qwen3.6-27B-DFlash")
+		spec := `{"method":"dflash","model":"z-lab/Qwen3.6-27B-DFlash"}`
+		out, err := resolveDraftModelPath(spec, modelsDir)
+		Expect(err).ToNot(HaveOccurred())
+		Expect(out).To(MatchJSON(`{"method":"dflash","model":"` + want + `"}`))
+	})
+
+	It("rewrites a models-dir-relative path", func() {
+		want := draftDir("drafts__dflash")
+		spec := `{"method":"dflash","model":"drafts__dflash"}`
+		out, err := resolveDraftModelPath(spec, modelsDir)
+		Expect(err).ToNot(HaveOccurred())
+		Expect(out).To(ContainSubstring(want))
+	})
+
+	It("leaves an absolute path that already resolves alone", func() {
+		abs := draftDir("elsewhere")
+		spec := `{"method":"dflash","model":"` + abs + `"}`
+		out, err := resolveDraftModelPath(spec, modelsDir)
+		Expect(err).ToNot(HaveOccurred())
+		Expect(out).To(MatchJSON(spec))
+	})
+
+	It("fails with an actionable error when the draft is nowhere on disk", func() {
+		// Silently passing the repo id through would surface as an HF-cache
+		// miss inside the engine, which reads as "your model is broken".
+		spec := `{"method":"dflash","model":"z-lab/Not-Downloaded"}`
+		_, err := resolveDraftModelPath(spec, modelsDir)
+		Expect(err).To(HaveOccurred())
+		Expect(err.Error()).To(ContainSubstring("z-lab/Not-Downloaded"))
+		Expect(err.Error()).To(ContainSubstring(modelsDir))
+	})
+
+	It("requires a model key for dflash", func() {
+		_, err := resolveDraftModelPath(`{"method":"dflash"}`, modelsDir)
+		Expect(err).To(HaveOccurred())
+		Expect(err.Error()).To(ContainSubstring("model"))
+	})
+
+	It("leaves mtp and ngram configs untouched", func() {
+		// Neither has a separate draft checkpoint to resolve.
+		for _, spec := range []string{
+			`{"method":"mtp"}`,
+			`{"method":"ngram","num_speculative_tokens":4}`,
+		} {
+			out, err := resolveDraftModelPath(spec, modelsDir)
+			Expect(err).ToNot(HaveOccurred())
+			Expect(out).To(MatchJSON(spec))
+		}
+	})
+
+	It("passes a malformed document through for the engine to reject", func() {
+		// The engine owns config validation and produces the better message.
+		out, err := resolveDraftModelPath(`{not json`, modelsDir)
+		Expect(err).ToNot(HaveOccurred())
+		Expect(out).To(Equal(`{not json`))
+	})
+
+	It("is a no-op on an empty config", func() {
+		out, err := resolveDraftModelPath("", modelsDir)
+		Expect(err).ToNot(HaveOccurred())
+		Expect(out).To(BeEmpty())
+	})
+})
+
 var _ = Describe("validModelPath", func() {
 	It("accepts a .gguf file", func() {
 		dir := GinkgoT().TempDir()

backend/go/whisper/Makefile

@@ -8,7 +8,7 @@ JOBS?=$(shell nproc --ignore=1)
 
 # whisper.cpp version
 WHISPER_REPO?=https://github.com/ggml-org/whisper.cpp
-WHISPER_CPP_VERSION?=2ca53bb45e38748d07b310eeb36245a7157ac882
+WHISPER_CPP_VERSION?=64d57d3df5c8dacee098577257edcaa154bf5ef3
 SO_TARGET?=libgowhisper.so
 
 CMAKE_ARGS+=-DBUILD_SHARED_LIBS=OFF

core/config/vllm_spec.go

@@ -0,0 +1,117 @@
+package config
+
+// Speculative-decoding auto-defaults for the vllm-cpp backend, the safetensors
+// counterpart of the GGUF/llama.cpp hook in mtp.go.
+//
+// The two engines detect and spell the same feature differently. llama.cpp
+// reads `<arch>.nextn_predict_layers` out of the GGUF header and takes
+// `spec_type:draft-mtp` in `options:`; vllm.cpp reads `mtp_num_hidden_layers`
+// out of the checkpoint's config.json and takes vLLM's own
+// `--speculative-config` JSON, which LocalAI carries in `engine_args`. The
+// engine resolves the draft depth and the default k itself, so the config only
+// has to name the method.
+
+import (
+	"encoding/json"
+
+	"github.com/mudler/xlog"
+)
+
+// hfSpecConfig is the subset of a HuggingFace config.json that decides whether
+// speculative decoding can be auto-enabled.
+type hfSpecConfig struct {
+	ModelType string `json:"model_type"`
+	// MtpNumHiddenLayers is the MTP head depth (upstream speculative.py reads
+	// it as n_predict for the qwen3_5 / qwen3_5_moe families).
+	MtpNumHiddenLayers uint32 `json:"mtp_num_hidden_layers"`
+	// DFlashConfig marks a z-lab DFlash DRAFT checkpoint (mask_token_id +
+	// target_layer_ids). Its presence means this repo is a draft, not a
+	// servable target.
+	DFlashConfig json.RawMessage `json:"dflash_config"`
+	// TextConfig is where multimodal checkpoints nest the language-model
+	// config, and therefore the MTP depth.
+	TextConfig *hfSpecConfig `json:"text_config"`
+}
+
+// parseHFSpecConfig decodes the speculative-relevant subset of a config.json.
+// A document that does not parse yields nothing rather than an error: detection
+// is best-effort and must never break an import.
+func parseHFSpecConfig(configJSON []byte) (hfSpecConfig, bool) {
+	if len(configJSON) == 0 {
+		return hfSpecConfig{}, false
+	}
+	var c hfSpecConfig
+	if err := json.Unmarshal(configJSON, &c); err != nil {
+		xlog.Debug("[vllm-spec] config.json did not parse; skipping detection", "error", err)
+		return hfSpecConfig{}, false
+	}
+	return c, true
+}
+
+// IsDFlashDraftConfig reports whether a HuggingFace config.json describes a
+// DFlash DRAFT checkpoint. Unlike MTP - whose head ships inside the target
+// checkpoint's `mtp.*` tensors - a DFlash draft is its own repo that can only
+// run paired with a target it verifies against, so it must never be configured
+// as a standalone model.
+func IsDFlashDraftConfig(configJSON []byte) bool {
+	c, ok := parseHFSpecConfig(configJSON)
+	if !ok {
+		return false
+	}
+	return len(c.DFlashConfig) > 0 ||
+		(c.TextConfig != nil && len(c.TextConfig.DFlashConfig) > 0)
+}
+
+// HasSafetensorsMTPHead reports whether a HuggingFace config.json declares a
+// self-speculating Multi-Token Prediction head, returning its depth. The depth
+// is informational: vllm.cpp resolves n_predict and the default
+// num_speculative_tokens from the checkpoint itself.
+//
+// DFlash drafts are excluded for the same reason `gemma4-assistant` GGUFs are
+// excluded from the llama.cpp hook: they carry head metadata but cannot
+// self-speculate.
+//
+// NOTE this is a safetensors-only signal. vllm.cpp rejects an MTP config over a
+// GGUF source, because the `mtp.*` draft tensors only exist in the safetensors
+// checkpoint - so the GGUF import path must not use this.
+func HasSafetensorsMTPHead(configJSON []byte) (uint32, bool) {
+	c, ok := parseHFSpecConfig(configJSON)
+	if !ok {
+		return 0, false
+	}
+	if IsDFlashDraftConfig(configJSON) {
+		return 0, false
+	}
+	n := c.MtpNumHiddenLayers
+	if n == 0 && c.TextConfig != nil {
+		n = c.TextConfig.MtpNumHiddenLayers
+	}
+	return n, n > 0
+}
+
+// ApplyVLLMSpeculativeDefaults enables MTP speculative decoding in cfg's
+// engine_args when nothing is configured there yet. It is a no-op when the user
+// already set a speculative_config, so an explicit choice (a different method,
+// an explicit k, a DFlash draft) is never clobbered.
+//
+// `layers` is the detected head depth and is only used for the diagnostic log
+// line - the engine derives the real k from the checkpoint.
+func ApplyVLLMSpeculativeDefaults(cfg *ModelConfig, layers uint32) {
+	if cfg == nil {
+		return
+	}
+	if _, set := cfg.EngineArgs["speculative_config"]; set {
+		xlog.Debug("[vllm-spec] MTP head detected but speculative_config already configured; leaving user choice intact",
+			"name", cfg.Name, "mtp_num_hidden_layers", layers)
+		return
+	}
+	if cfg.EngineArgs == nil {
+		cfg.EngineArgs = map[string]any{}
+	}
+	// Only the method: vllm.cpp defaults num_speculative_tokens to the
+	// checkpoint's own n_predict (speculative.py:865-875), which is the right
+	// value far more reliably than anything guessable here.
+	cfg.EngineArgs["speculative_config"] = map[string]any{"method": "mtp"}
+	xlog.Info("[vllm-spec] MTP head detected; enabling mtp speculative decoding",
+		"name", cfg.Name, "mtp_num_hidden_layers", layers)
+}

core/config/vllm_spec_test.go

@@ -0,0 +1,117 @@
+package config_test
+
+import (
+	. "github.com/mudler/LocalAI/core/config"
+
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+)
+
+var _ = Describe("vllm-cpp speculative-decoding auto-defaults", func() {
+	Context("HasSafetensorsMTPHead", func() {
+		It("detects a top-level mtp_num_hidden_layers", func() {
+			n, ok := HasSafetensorsMTPHead([]byte(`{
+				"model_type": "qwen3_5_moe",
+				"mtp_num_hidden_layers": 1
+			}`))
+			Expect(ok).To(BeTrue())
+			Expect(n).To(Equal(uint32(1)))
+		})
+
+		It("detects the head nested under text_config", func() {
+			// Multimodal checkpoints nest the language-model config, which is
+			// where the MTP depth lives (mirrors the engine's own resolution
+			// off config.raw text_config).
+			n, ok := HasSafetensorsMTPHead([]byte(`{
+				"model_type": "qwen3_5_moe",
+				"text_config": {"mtp_num_hidden_layers": 2}
+			}`))
+			Expect(ok).To(BeTrue())
+			Expect(n).To(Equal(uint32(2)))
+		})
+
+		It("reports no head when the key is absent", func() {
+			n, ok := HasSafetensorsMTPHead([]byte(`{"model_type": "llama"}`))
+			Expect(ok).To(BeFalse())
+			Expect(n).To(BeZero())
+		})
+
+		It("reports no head for a zero depth", func() {
+			_, ok := HasSafetensorsMTPHead([]byte(`{"mtp_num_hidden_layers": 0}`))
+			Expect(ok).To(BeFalse())
+		})
+
+		It("ignores a DFlash draft checkpoint", func() {
+			// A DFlash draft is a SEPARATE checkpoint that cannot serve alone:
+			// it needs a target to verify against. Same exclusion the GGUF path
+			// makes for gemma4-assistant drafts.
+			_, ok := HasSafetensorsMTPHead([]byte(`{
+				"model_type": "qwen3_dflash",
+				"mtp_num_hidden_layers": 1,
+				"dflash_config": {"mask_token_id": 151666, "target_layer_ids": [0, 1]}
+			}`))
+			Expect(ok).To(BeFalse())
+		})
+
+		It("reports no head on unparseable JSON", func() {
+			_, ok := HasSafetensorsMTPHead([]byte(`{not json`))
+			Expect(ok).To(BeFalse())
+		})
+
+		It("reports no head on empty input", func() {
+			_, ok := HasSafetensorsMTPHead(nil)
+			Expect(ok).To(BeFalse())
+		})
+	})
+
+	Context("IsDFlashDraftConfig", func() {
+		It("recognises a draft by its dflash_config block", func() {
+			Expect(IsDFlashDraftConfig([]byte(`{
+				"dflash_config": {"mask_token_id": 151666, "target_layer_ids": [0]}
+			}`))).To(BeTrue())
+		})
+
+		It("does not flag an ordinary checkpoint", func() {
+			Expect(IsDFlashDraftConfig([]byte(`{"model_type": "qwen3_5_moe"}`))).To(BeFalse())
+		})
+	})
+
+	Context("ApplyVLLMSpeculativeDefaults", func() {
+		It("writes the mtp method into engine_args", func() {
+			cfg := &ModelConfig{Name: "qwen"}
+			ApplyVLLMSpeculativeDefaults(cfg, 1)
+			Expect(cfg.EngineArgs).To(HaveKey("speculative_config"))
+			spec, ok := cfg.EngineArgs["speculative_config"].(map[string]any)
+			Expect(ok).To(BeTrue())
+			Expect(spec["method"]).To(Equal("mtp"))
+		})
+
+		It("leaves an existing speculative_config alone", func() {
+			cfg := &ModelConfig{
+				Name: "qwen",
+				LLMConfig: LLMConfig{
+					EngineArgs: map[string]any{
+						"speculative_config": map[string]any{"method": "ngram", "num_speculative_tokens": 4},
+					},
+				},
+			}
+			ApplyVLLMSpeculativeDefaults(cfg, 1)
+			spec := cfg.EngineArgs["speculative_config"].(map[string]any)
+			Expect(spec["method"]).To(Equal("ngram"))
+		})
+
+		It("preserves unrelated engine_args keys", func() {
+			cfg := &ModelConfig{
+				Name:      "qwen",
+				LLMConfig: LLMConfig{EngineArgs: map[string]any{"max_num_seqs": 32}},
+			}
+			ApplyVLLMSpeculativeDefaults(cfg, 1)
+			Expect(cfg.EngineArgs).To(HaveKeyWithValue("max_num_seqs", 32))
+			Expect(cfg.EngineArgs).To(HaveKey("speculative_config"))
+		})
+
+		It("tolerates a nil config", func() {
+			Expect(func() { ApplyVLLMSpeculativeDefaults(nil, 1) }).ToNot(Panic())
+		})
+	})
+})

core/gallery/importers/llama-cpp.go

@@ -298,7 +298,15 @@ func (i *LlamaCPPImporter) Import(details Details) (gallery.ModelConfig, error)
 	// imported configs already carry spec_type:draft-mtp before the model is
 	// ever loaded - users see it in the YAML preview rather than discovering
 	// it after the first start.
-	maybeApplyMTPDefaults(&modelConfig, details, &cfg)
+	//
+	// vllm-cpp is excluded on both counts: `spec_type:*` are llama.cpp option
+	// keys it does not read, and vllm.cpp rejects an MTP config over a GGUF
+	// source outright (the `mtp.*` draft tensors exist only in the safetensors
+	// checkpoint). Its MTP auto-config runs in the vllm importer instead, over
+	// the safetensors config.json.
+	if backend != "vllm-cpp" {
+		maybeApplyMTPDefaults(&modelConfig, details, &cfg)
+	}
 
 	data, err := yaml.Marshal(modelConfig)
 	if err != nil {

core/gallery/importers/vllm.go

@@ -1,13 +1,21 @@
 package importers
 
 import (
+	"context"
 	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
 	"path/filepath"
 	"strings"
+	"time"
 
 	"github.com/mudler/LocalAI/core/config"
 	"github.com/mudler/LocalAI/core/gallery"
 	"github.com/mudler/LocalAI/core/schema"
+	"github.com/mudler/LocalAI/pkg/downloader"
+	"github.com/mudler/LocalAI/pkg/httpclient"
+	"github.com/mudler/xlog"
 	"go.yaml.in/yaml/v2"
 )
 
@@ -107,6 +115,12 @@ func (i *VLLMImporter) Import(details Details) (gallery.ModelConfig, error) {
 		// vllm python backend, so use_tokenizer_template carries over), but
 		// tool/reasoning parsing is the engine's own autoparser pipeline -
 		// the vllm-python tool_parser/reasoning_parser options don't apply.
+		//
+		// Auto-detect a Multi-Token Prediction head, the safetensors analogue
+		// of the llama-cpp importer's GGUF hook, so a freshly imported
+		// Qwen3.5 / Qwen3.6 config already carries speculative decoding in its
+		// engine_args instead of leaving the throughput on the table.
+		maybeApplyVLLMSpeculativeDefaults(&modelConfig, details)
 	} else {
 		// Auto-detect tool_parser and reasoning_parser for known model families.
 		// Surfacing them in the generated YAML lets users see and edit the choices.
@@ -132,3 +146,89 @@ func (i *VLLMImporter) Import(details Details) (gallery.ModelConfig, error) {
 		ConfigFile:  string(data),
 	}, nil
 }
+
+// maxSpecConfigProbeBytes caps the config.json body we read. Real ones are a
+// few KB; the cap keeps a hostile or mislabelled URL from streaming into the
+// importer.
+const maxSpecConfigProbeBytes = 1 << 20 // 1 MiB
+
+// specConfigProbeTimeout bounds the config.json fetch. Detection is an
+// optimisation, so it must never hold an import open for long.
+const specConfigProbeTimeout = 30 * time.Second
+
+// specConfigFetcher is the seam the config.json probe goes through, so tests can
+// drive the whole import path without a network round trip.
+var specConfigFetcher = fetchProbeBody
+
+// maybeApplyVLLMSpeculativeDefaults fetches the repository's config.json and,
+// when it declares a Multi-Token Prediction head, enables MTP speculative
+// decoding in the emitted engine_args. This is the safetensors counterpart of
+// the llama-cpp importer's GGUF header probe.
+//
+// Every failure is non-fatal and logged at debug: a network blip, a private
+// repo, or a config.json this doesn't understand must leave the import working
+// exactly as it did before, just without the speculative default.
+func maybeApplyVLLMSpeculativeDefaults(modelConfig *config.ModelConfig, details Details) {
+	probeURL := vllmSpecProbeURL(details)
+	if probeURL == "" {
+		return
+	}
+
+	body, err := specConfigFetcher(probeURL)
+	if err != nil {
+		xlog.Debug("[vllm-spec-importer] could not read config.json for MTP detection", "uri", probeURL, "error", err)
+		return
+	}
+
+	applySpecFromConfigJSON(modelConfig, body, details.URI)
+}
+
+// applySpecFromConfigJSON is the decision half of the probe, split out so it can
+// be exercised without a network round trip.
+func applySpecFromConfigJSON(modelConfig *config.ModelConfig, body []byte, uri string) {
+	if config.IsDFlashDraftConfig(body) {
+		// A DFlash draft cannot serve on its own - it only proposes tokens for
+		// a target model to verify. Say so rather than emitting a config that
+		// would fail at load.
+		xlog.Warn("[vllm-spec-importer] this repository is a DFlash DRAFT checkpoint, not a servable model; "+
+			"import the TARGET model and point engine_args.speculative_config at this repo "+
+			`({"method":"dflash","model":"<this repo>"})`, "uri", uri)
+		return
+	}
+
+	n, ok := config.HasSafetensorsMTPHead(body)
+	if !ok {
+		return
+	}
+	config.ApplyVLLMSpeculativeDefaults(modelConfig, n)
+}
+
+// vllmSpecProbeURL returns the HTTP(S) URL of the repository's config.json, or
+// "" when the import isn't backed by a HuggingFace repo we can fetch from (a
+// local directory import, an OCI artifact, ...).
+func vllmSpecProbeURL(details Details) string {
+	if details.HuggingFace == nil || details.HuggingFace.ModelID == "" {
+		return ""
+	}
+	return resolveHTTPProbe(downloader.HuggingFacePrefix + details.HuggingFace.ModelID + "/config.json")
+}
+
+// fetchProbeBody GETs a small remote JSON document under a short timeout.
+func fetchProbeBody(url string) ([]byte, error) {
+	ctx, cancel := context.WithTimeout(context.Background(), specConfigProbeTimeout)
+	defer cancel()
+
+	req, err := http.NewRequestWithContext(ctx, http.MethodGet, url, nil)
+	if err != nil {
+		return nil, err
+	}
+	resp, err := httpclient.NewWithTimeout(specConfigProbeTimeout).Do(req)
+	if err != nil {
+		return nil, err
+	}
+	defer func() { _ = resp.Body.Close() }()
+	if resp.StatusCode != http.StatusOK {
+		return nil, fmt.Errorf("unexpected status %d", resp.StatusCode)
+	}
+	return io.ReadAll(io.LimitReader(resp.Body, maxSpecConfigProbeBytes))
+}

core/gallery/importers/vllm_spec_internal_test.go

@@ -0,0 +1,118 @@
+package importers
+
+import (
+	"encoding/json"
+	"errors"
+
+	"github.com/mudler/LocalAI/core/config"
+	hfapi "github.com/mudler/LocalAI/pkg/huggingface-api"
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+)
+
+var _ = Describe("vllm-cpp speculative auto-config (importer)", func() {
+	Context("applySpecFromConfigJSON", func() {
+		It("enables mtp when the checkpoint declares an MTP head", func() {
+			cfg := &config.ModelConfig{Name: "qwen3.5"}
+			applySpecFromConfigJSON(cfg, []byte(`{
+				"model_type": "qwen3_5_moe",
+				"mtp_num_hidden_layers": 1
+			}`), "huggingface://Qwen/Qwen3.5-A3B")
+			Expect(cfg.EngineArgs).To(HaveKeyWithValue("speculative_config",
+				map[string]any{"method": "mtp"}))
+		})
+
+		It("leaves a plain checkpoint untouched", func() {
+			cfg := &config.ModelConfig{Name: "llama"}
+			applySpecFromConfigJSON(cfg, []byte(`{"model_type": "llama"}`), "huggingface://meta/llama")
+			Expect(cfg.EngineArgs).To(BeEmpty())
+		})
+
+		It("refuses to configure a DFlash draft as a servable model", func() {
+			// The draft only proposes tokens; configuring it standalone would
+			// produce a model that cannot load.
+			cfg := &config.ModelConfig{Name: "dflash-draft"}
+			applySpecFromConfigJSON(cfg, []byte(`{
+				"model_type": "qwen3_dflash",
+				"dflash_config": {"mask_token_id": 151666, "target_layer_ids": [0, 1]}
+			}`), "huggingface://z-lab/Qwen3.6-27B-DFlash")
+			Expect(cfg.EngineArgs).To(BeEmpty())
+		})
+
+		It("survives a config.json it cannot parse", func() {
+			cfg := &config.ModelConfig{Name: "weird"}
+			Expect(func() {
+				applySpecFromConfigJSON(cfg, []byte(`<html>404</html>`), "huggingface://a/b")
+			}).ToNot(Panic())
+			Expect(cfg.EngineArgs).To(BeEmpty())
+		})
+	})
+
+	Context("Import over a repository with an MTP head", func() {
+		var restore func()
+
+		BeforeEach(func() {
+			original := specConfigFetcher
+			restore = func() { specConfigFetcher = original }
+		})
+		AfterEach(func() { restore() })
+
+		importWith := func(backend, configJSON string) string {
+			specConfigFetcher = func(string) ([]byte, error) {
+				return []byte(configJSON), nil
+			}
+			importer := &VLLMImporter{}
+			out, err := importer.Import(Details{
+				URI:         "huggingface://Qwen/Qwen3.5-A3B",
+				Preferences: json.RawMessage(`{"backend": "` + backend + `"}`),
+				HuggingFace: &hfapi.ModelDetails{ModelID: "Qwen/Qwen3.5-A3B"},
+			})
+			Expect(err).ToNot(HaveOccurred())
+			return out.ConfigFile
+		}
+
+		It("emits engine_args.speculative_config for vllm-cpp", func() {
+			yaml := importWith("vllm-cpp", `{"model_type":"qwen3_5_moe","mtp_num_hidden_layers":1}`)
+			Expect(yaml).To(ContainSubstring("engine_args:"))
+			Expect(yaml).To(ContainSubstring("speculative_config:"))
+			Expect(yaml).To(ContainSubstring("method: mtp"))
+		})
+
+		It("emits nothing speculative for the python vllm backend", func() {
+			// The python backend has its own speculative surface and its own
+			// version-dependent MTP support; this hook is vllm-cpp only.
+			yaml := importWith("vllm", `{"model_type":"qwen3_5_moe","mtp_num_hidden_layers":1}`)
+			Expect(yaml).NotTo(ContainSubstring("speculative_config"))
+		})
+
+		It("emits nothing speculative when the probe fails", func() {
+			specConfigFetcher = func(string) ([]byte, error) {
+				return nil, errors.New("network down")
+			}
+			importer := &VLLMImporter{}
+			out, err := importer.Import(Details{
+				URI:         "huggingface://Qwen/Qwen3.5-A3B",
+				Preferences: json.RawMessage(`{"backend": "vllm-cpp"}`),
+				HuggingFace: &hfapi.ModelDetails{ModelID: "Qwen/Qwen3.5-A3B"},
+			})
+			Expect(err).ToNot(HaveOccurred())
+			Expect(out.ConfigFile).NotTo(ContainSubstring("speculative_config"))
+		})
+	})
+
+	Context("vllmSpecProbeURL", func() {
+		It("resolves the repository's config.json to an HTTPS URL", func() {
+			url := vllmSpecProbeURL(Details{
+				URI:         "huggingface://Qwen/Qwen3.5-A3B",
+				HuggingFace: &hfapi.ModelDetails{ModelID: "Qwen/Qwen3.5-A3B"},
+			})
+			Expect(url).To(ContainSubstring("Qwen/Qwen3.5-A3B"))
+			Expect(url).To(HaveSuffix("config.json"))
+			Expect(url).To(HavePrefix("https://"))
+		})
+
+		It("skips the probe when there is no HuggingFace repo behind the import", func() {
+			Expect(vllmSpecProbeURL(Details{URI: "/models/local-dir"})).To(BeEmpty())
+		})
+	})
+})

core/http/endpoints/mcp/localai_assistant_test.go

@@ -154,22 +154,6 @@ func (stubClient) GetRouterDecisions(_ context.Context, _ localaitools.RouterDec
 	return []localaitools.RouterDecision{}, nil
 }
 
-func (stubClient) ListScheduling(_ context.Context) ([]localaitools.ModelSchedulingConfig, error) {
-	return []localaitools.ModelSchedulingConfig{}, nil
-}
-
-func (stubClient) GetScheduling(_ context.Context, _ string) (*localaitools.ModelSchedulingConfig, error) {
-	return &localaitools.ModelSchedulingConfig{}, nil
-}
-
-func (stubClient) SetScheduling(_ context.Context, _ localaitools.SetSchedulingRequest) (*localaitools.ModelSchedulingConfig, error) {
-	return &localaitools.ModelSchedulingConfig{}, nil
-}
-
-func (stubClient) DeleteScheduling(_ context.Context, _ string) error {
-	return nil
-}
-
 var _ = Describe("LocalAIAssistantHolder", func() {
 	var ctx context.Context
 

core/http/react-ui/package-lock.json

@@ -21,9 +21,10 @@
         "@fortawesome/fontawesome-free": "^6.7.2",
         "@lezer/highlight": "^1.2.1",
         "@modelcontextprotocol/ext-apps": "^1.2.2",
-        "@modelcontextprotocol/sdk": "^1.25.1",
+        "@modelcontextprotocol/sdk": "^1.30.0",
         "dompurify": "^3.4.12",
         "highlight.js": "^11.11.1",
+        "hono": "4.12.25",
         "i18next": "^26.0.8",
         "i18next-browser-languagedetector": "^8.2.1",
         "i18next-http-backend": "^3.0.6",
@@ -944,11 +945,12 @@
       }
     },
     "node_modules/@modelcontextprotocol/sdk": {
-      "version": "1.27.1",
-      "resolved": "https://registry.npmjs.org/@modelcontextprotocol/sdk/-/sdk-1.27.1.tgz",
-      "integrity": "sha512-sr6GbP+4edBwFndLbM60gf07z0FQ79gaExpnsjMGePXqFcSSb7t6iscpjk9DhFhwd+mTEQrzNafGP8/iGGFYaA==",
+      "version": "1.30.0",
+      "resolved": "https://registry.npmjs.org/@modelcontextprotocol/sdk/-/sdk-1.30.0.tgz",
+      "integrity": "sha512-xKd8OIzlqNzcqcNumGAa6g+PW2kjD5vrpcKOnfldAUPP3j7lnqMPwlTXQm8gF+UwH72z0lqaRbjr9hqGz0eITA==",
+      "license": "MIT",
       "dependencies": {
-        "@hono/node-server": "^1.19.9",
+        "@hono/node-server": "^1.19.9 || ^2.0.5",
         "ajv": "^8.17.1",
         "ajv-formats": "^3.0.1",
         "content-type": "^1.0.5",
@@ -1718,10 +1720,11 @@
       "dev": true
     },
     "node_modules/brace-expansion": {
-      "version": "1.1.12",
-      "resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-1.1.12.tgz",
-      "integrity": "sha512-9T9UjW3r0UW5c1Q7GTwllptXwhvYmEzFhzMfZ9H7FQWt+uZePjZPjBP/W1ZEyZ1twGWom5/56TF4lPcqjnDHcg==",
+      "version": "1.1.18",
+      "resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-1.1.18.tgz",
+      "integrity": "sha512-Edep/X9fGqVNmzKBVsDYIOtD+z1tuezV70LBjdCst9Tqu76lsnvRiZ6oTic1n+/BIwX6QDGAO94PN4N2SADvtw==",
       "dev": true,
+      "license": "MIT",
       "dependencies": {
         "balanced-match": "^1.0.0",
         "concat-map": "0.0.1"
@@ -3432,9 +3435,9 @@
       }
     },
     "node_modules/hono": {
-      "version": "4.12.31",
-      "resolved": "https://registry.npmjs.org/hono/-/hono-4.12.31.tgz",
-      "integrity": "sha512-zJIHFrl6bq3RDd2YusFNCDlM8qUprxKswyi/OPzPyzKDdyBXDqWx8bZlZ7R+saTdSTatUmb3O7K4SspGPaEOQg==",
+      "version": "4.12.25",
+      "resolved": "https://registry.npmjs.org/hono/-/hono-4.12.25.tgz",
+      "integrity": "sha512-2NFaIyNVgJmBs/ecmtGzlmluTFs5cHEWGTdu0t1HBwYzoGXOL5nUQBRMXsXWla5i4KkG//QMzVP88m1+I3fdAQ==",
       "license": "MIT",
       "engines": {
         "node": ">=16.9.0"
@@ -4383,16 +4386,16 @@
       }
     },
     "node_modules/istanbul-lib-processinfo/node_modules/brace-expansion": {
-      "version": "5.0.6",
-      "resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-5.0.6.tgz",
-      "integrity": "sha512-kLpxurY4Z4r9sgMsyG0Z9uzsBlgiU/EFKhj/h91/8yHu0edo7XuixOIH3VcJ8kkxs6/jPzoI6U9Vj3WqbMQ94g==",
+      "version": "5.0.9",
+      "resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-5.0.9.tgz",
+      "integrity": "sha512-ScQ4IuvIEF1TMlP7Zt+vjJ//9zlPb2SDcxWxM3bk8s6t6GGdJ7KO1dCcTidOPJKePW30LE/2cT7wCyPho9/Wxg==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
         "balanced-match": "^4.0.2"
       },
       "engines": {
-        "node": "18 || 20 || >=22"
+        "node": "20 || >=22"
       }
     },
     "node_modules/istanbul-lib-processinfo/node_modules/glob": {
@@ -5278,16 +5281,16 @@
       }
     },
     "node_modules/nyc/node_modules/brace-expansion": {
-      "version": "5.0.6",
-      "resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-5.0.6.tgz",
-      "integrity": "sha512-kLpxurY4Z4r9sgMsyG0Z9uzsBlgiU/EFKhj/h91/8yHu0edo7XuixOIH3VcJ8kkxs6/jPzoI6U9Vj3WqbMQ94g==",
+      "version": "5.0.9",
+      "resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-5.0.9.tgz",
+      "integrity": "sha512-ScQ4IuvIEF1TMlP7Zt+vjJ//9zlPb2SDcxWxM3bk8s6t6GGdJ7KO1dCcTidOPJKePW30LE/2cT7wCyPho9/Wxg==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
         "balanced-match": "^4.0.2"
       },
       "engines": {
-        "node": "18 || 20 || >=22"
+        "node": "20 || >=22"
       }
     },
     "node_modules/nyc/node_modules/convert-source-map": {
@@ -5974,10 +5977,11 @@
       }
     },
     "node_modules/quick-temp/node_modules/brace-expansion": {
-      "version": "2.1.0",
-      "resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-2.1.0.tgz",
-      "integrity": "sha512-TN1kCZAgdgweJhWWpgKYrQaMNHcDULHkWwQIspdtjV4Y5aurRdZpjAqn6yX3FPqTA9ngHCc4hJxMAMgGfve85w==",
+      "version": "2.1.4",
+      "resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-2.1.4.tgz",
+      "integrity": "sha512-hGfVzPxthbf3+2yjg/RBs60cB0FhqBS/zvdV/4wn4/BmN0bNMMHPc4V/BbFieqf1TKAGGAHnY4eSjajCl0f2Xg==",
       "dev": true,
+      "license": "MIT",
       "dependencies": {
         "balanced-match": "^1.0.0"
       }
@@ -6569,16 +6573,16 @@
       }
     },
     "node_modules/spawn-wrap/node_modules/brace-expansion": {
-      "version": "5.0.6",
-      "resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-5.0.6.tgz",
-      "integrity": "sha512-kLpxurY4Z4r9sgMsyG0Z9uzsBlgiU/EFKhj/h91/8yHu0edo7XuixOIH3VcJ8kkxs6/jPzoI6U9Vj3WqbMQ94g==",
+      "version": "5.0.9",
+      "resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-5.0.9.tgz",
+      "integrity": "sha512-ScQ4IuvIEF1TMlP7Zt+vjJ//9zlPb2SDcxWxM3bk8s6t6GGdJ7KO1dCcTidOPJKePW30LE/2cT7wCyPho9/Wxg==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
         "balanced-match": "^4.0.2"
       },
       "engines": {
-        "node": "18 || 20 || >=22"
+        "node": "20 || >=22"
       }
     },
     "node_modules/spawn-wrap/node_modules/foreground-child": {
@@ -6902,16 +6906,16 @@
       }
     },
     "node_modules/test-exclude/node_modules/brace-expansion": {
-      "version": "5.0.6",
-      "resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-5.0.6.tgz",
-      "integrity": "sha512-kLpxurY4Z4r9sgMsyG0Z9uzsBlgiU/EFKhj/h91/8yHu0edo7XuixOIH3VcJ8kkxs6/jPzoI6U9Vj3WqbMQ94g==",
+      "version": "5.0.9",
+      "resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-5.0.9.tgz",
+      "integrity": "sha512-ScQ4IuvIEF1TMlP7Zt+vjJ//9zlPb2SDcxWxM3bk8s6t6GGdJ7KO1dCcTidOPJKePW30LE/2cT7wCyPho9/Wxg==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
         "balanced-match": "^4.0.2"
       },
       "engines": {
-        "node": "18 || 20 || >=22"
+        "node": "20 || >=22"
       }
     },
     "node_modules/test-exclude/node_modules/glob": {

core/http/react-ui/package.json

@@ -35,7 +35,7 @@
     "@fortawesome/fontawesome-free": "^6.7.2",
     "@lezer/highlight": "^1.2.1",
     "@modelcontextprotocol/ext-apps": "^1.2.2",
-    "@modelcontextprotocol/sdk": "^1.25.1",
+    "@modelcontextprotocol/sdk": "^1.30.0",
     "dompurify": "^3.4.12",
     "highlight.js": "^11.11.1",
     "hono": "4.12.25",

docs/content/features/text-generation.md

@@ -918,6 +918,200 @@ options:
 The full list of registered parsers lives in `sglang.srt.function_call`
 and `sglang.srt.parser.reasoning_parser`.
 
+### vllm.cpp
+
+[vllm.cpp](https://github.com/mudler/vllm.cpp) is the LocalAI team's C++ port of
+vLLM: the same continuous-batching scheduler, paged KV cache and prefix caching,
+with no Python at inference time. It consumes either a HuggingFace safetensors
+model directory or a `.gguf` file, and applies the model's chat template,
+tool-call parsing and reasoning split engine-side.
+
+#### Setup
+
+```yaml
+name: vllm-cpp
+backend: vllm-cpp
+parameters:
+  model: "Qwen/Qwen3-4B"
+context_size: 8192
+template:
+  use_tokenizer_template: true
+```
+
+#### Configuring the engine with `engine_args`
+
+The same `engine_args:` map the vLLM and SGLang backends accept is honoured
+here, with keys spelled exactly as vLLM's own CLI flags - so a `speculative_config`
+or `kv_transfer_config` block written for vLLM works verbatim. Unknown keys are
+ignored rather than fatal; the engine validates the documents it is handed and
+reports a precise error at load.
+
+```yaml
+name: qwen35-a3b
+backend: vllm-cpp
+parameters:
+  model: "Qwen/Qwen3.5-A3B"
+context_size: 16384
+template:
+  use_tokenizer_template: true
+engine_args:
+  # KV cache sizing: num_blocks * block_size tokens of cache.
+  block_size: 32
+  num_blocks: 1024
+  # Concurrency and the per-step chunked-prefill token budget.
+  max_num_seqs: 32
+  max_num_batched_tokens: 8192
+  # Automatic prefix caching. Omit to keep the model's own default
+  # (on for dense models, off for hybrid / attention-free ones).
+  enable_prefix_caching: true
+  # Scheduler admission order: fcfs (default), priority, or lpm
+  # (cache-aware longest-prefix-match; needs prefix caching to have any effect).
+  scheduling_policy: lpm
+```
+
+| Key | Meaning | Default |
+|-----|---------|---------|
+| `block_size` | KV-cache block size, in tokens per block | 32 |
+| `num_blocks` | KV-cache blocks to allocate | 256 |
+| `max_model_len` | Max sequence length; also settable as `context_size` / `max_model_len` | model config |
+| `max_num_seqs` | Max concurrent sequences the scheduler admits | 8 |
+| `max_num_batched_tokens` | Per-step chunked-prefill token budget | per-arch (2048 dense, 4096/8192 MoE) |
+| `enable_prefix_caching` | Automatic prefix caching; `enable_radix_attention` is an accepted alias | model default |
+| `enable_jump_forward` | Jump-forward decoding, which emits grammar-forced tokens without a model step. Only affects constrained requests (`grammar`, JSON schema) | off |
+| `scheduling_policy` | `fcfs`, `priority`, or `lpm` | `fcfs` |
+| `tool_parser` / `reasoning_parser` | Force a parser instead of chat-template auto-detection | auto |
+| `tokenizer_config` | Override the `tokenizer_config.json` the chat template is read from | `<model_dir>/tokenizer_config.json` |
+| `speculative_config` | Speculative decoding (see below) | disabled |
+| `kv_transfer_config` | External KV connector / LMCache (see below) | none |
+
+Raising `max_num_batched_tokens` lets more prefill land in a single step, at the
+cost of decode latency for requests queued behind it. The default deliberately
+does not scale with `max_num_seqs`, which is what keeps a large concurrent
+prefill from blowing up the per-step activation on the hybrid architectures.
+
+`enable_prefix_caching` and `enable_jump_forward` are tri-state at the engine
+boundary: omitting the key defers to a default (the model's own capability for
+prefix caching, an environment variable for jump forward), while an explicit
+`false` forces the feature off. Those are genuinely different - prefix caching
+defaults *on* for dense models - so write the key only when you mean to override.
+
+#### Speculative decoding
+
+`speculative_config:` takes the same JSON object as vLLM's
+`--speculative-config`. Three methods are supported.
+
+> **Architecture limit.** At the current engine pin, `mtp` and `dflash` are
+> **Qwen3.5 / Qwen3.6 only**. The engine builds a widened speculative KV cache
+> directly for those families rather than through the model registry, so a
+> speculative config on any other architecture (Llama, GLM, Gemma, Mistral, ...)
+> will not work regardless of checkpoint format. `ngram` needs no draft weights
+> and is not subject to this limit.
+
+> **Format support.** `mtp` and `dflash` now work from a `.gguf` target as well
+> as safetensors. An MTP head is read from the GGUF's `nextn.*` tensors when the
+> file declares `<arch>.nextn_predict_layers`; a GGUF exported WITHOUT the head
+> (converted with `--no-mtp`, or predating llama.cpp's Qwen3.5 MTP support) is
+> refused at load naming that as the reason. A DFlash draft may itself be a
+> `dflash`-arch GGUF, and the target may be a GGUF too. `ngram` needs no draft
+> weights and works on any format.
+
+**MTP** (Multi-Token Prediction) uses a draft head shipped inside the target
+checkpoint's own `mtp.*` tensors, so there is no second model to download. It
+requires a **safetensors** checkpoint - the `mtp.*` tensors do not survive GGUF
+conversion, and an MTP config over a `.gguf` model is rejected at load.
+
+```yaml
+engine_args:
+  speculative_config:
+    method: mtp
+    # Optional; defaults to the checkpoint's own head depth, which is
+    # usually the right value. Must be a multiple of that depth.
+    num_speculative_tokens: 1
+```
+
+**DFlash** uses a separate block-diffusion drafter that proposes a whole block
+of tokens in one non-autoregressive forward pass. Unlike MTP, the draft is its
+own checkpoint, so `model:` is **required**:
+
+```yaml
+engine_args:
+  speculative_config:
+    method: dflash
+    model: z-lab/Qwen3.6-27B-DFlash
+    num_speculative_tokens: 4
+```
+
+The draft shares the *target's* `embed_tokens` and `lm_head`, so both must come
+from the same model family and the target must be safetensors.
+
+**The engine does not download the draft.** `model:` is resolved, in order,
+as a path as given, then as the last path segment under LocalAI's models
+directory (`z-lab/Qwen3.6-27B-DFlash` → `<models>/Qwen3.6-27B-DFlash`, which is
+what LocalAI's own downloader produces), then as the whole reference under the
+models directory. Install the draft into LocalAI first, or give an absolute path
+to a directory containing `config.json`. If none of those resolve, the load
+fails immediately naming every location that was tried, rather than reporting a
+missing checkpoint from inside the engine.
+
+**N-gram** needs no draft model at all - it proposes from the prompt's own
+suffix history. `num_speculative_tokens` is required:
+
+```yaml
+engine_args:
+  speculative_config:
+    method: ngram
+    num_speculative_tokens: 4
+    prompt_lookup_min: 5
+    prompt_lookup_max: 5
+```
+
+> **Auto-configuration on import.** When you import a safetensors repository
+> with `backend: vllm-cpp`, LocalAI reads the checkpoint's `config.json` and, if
+> it declares an MTP head (`mtp_num_hidden_layers`), writes
+> `speculative_config: {method: mtp}` into the generated `engine_args` for you.
+> An explicit `speculative_config` in your own config is never overwritten.
+> Importing a DFlash *draft* repository is refused with a warning: a drafter
+> cannot serve on its own, so import the target model and point
+> `speculative_config.model` at the draft.
+
+#### External KV cache with LMCache
+
+`kv_transfer_config:` takes vLLM's `--kv-transfer-config` JSON and selects an
+external KV-cache connector. The `lm://` LMCache client lets prefill KV be
+stored to and reloaded from a shared `lmcache.v1.server`, so a prefix computed
+by one replica does not have to be recomputed by the next:
+
+```yaml
+engine_args:
+  kv_transfer_config:
+    kv_connector: LMCacheConnector
+    kv_role: kv_both          # required whenever kv_connector is set
+    kv_connector_extra_config:
+      host: 127.0.0.1
+      port: 65432
+```
+
+`kv_role` is one of `kv_producer` (store only), `kv_consumer` (load only), or
+`kv_both`. An unregistered connector name, a missing role, or a malformed
+document fails the load with an explicit error rather than silently running
+without the cache.
+
+#### Legacy `options:` list
+
+Earlier versions configured this backend through the flat `options:` list, and
+those configs keep working. Every key in the table above is still read from
+there in `key:value` form, and `engine_args` wins on any key set in both:
+
+```yaml
+options:
+  - max_num_seqs:32
+  - enable_prefix_caching:true
+```
+
+New configs should prefer `engine_args:`, which is the only place the nested
+`speculative_config` / `kv_transfer_config` documents can be written naturally
+rather than as a single-line JSON string.
+
 ### Transformers
 
 [Transformers](https://huggingface.co/docs/transformers/index) is a State-of-the-art Machine Learning library for PyTorch, TensorFlow, and JAX.

docs/content/whats-new.md

@@ -6,27 +6,11 @@ url = '/basics/news/'
 icon = "newspaper"
 +++
 
-Release notes have been now moved completely over Github releases.
+LocalAI news is published in two places, both kept current:
 
-You can see the release notes [here](https://github.com/mudler/LocalAI/releases).
+- **[Blog](https://localai.io/blog/)** for release write-ups, benchmark reports and engineering notes.
+- **[GitHub Releases](https://github.com/mudler/LocalAI/releases)** for the full changelog of every version.
 
-## 2026 Highlights
+For how the project got here, read [LocalAI, from March 2023 to now](https://localai.io/blog/localai-since-march-2023/).
 
-- **August 2026**: [Text moderation](/features/moderation/) - new OpenAI-compatible `POST /v1/moderations` endpoint. It uses any local completion model with a constrained JSON grammar and returns the standard safety categories, scores, and per-input flags.
-- **July 2026**: [LongCat video and avatar generation](/features/video-generation/) - dedicated CUDA backend for `LongCat-Video` text/image-to-video and `LongCat-Video-Avatar-1.5` speech-driven avatars. Includes multi-segment continuation, portrait and recorded-audio inputs in Studio, and an SDPA CUDA 13 ARM64 build for DGX Spark.
-- **April 2026**: [Audio Transform](/features/audio-transform/) - generic audio-in / audio-out endpoint with optional reference signal. First implementation: [LocalVQE](https://github.com/localai-org/LocalVQE) C++ backend (joint AEC + noise suppression + dereverberation, DeepVQE-style). Both batch (`POST /audio/transformations`) and bidirectional WebSocket streaming (`/audio/transformations/stream`). Studio "Transform" tab with synchronized waveform players for input / reference / output.
-- **April 2026**: [Face recognition backend](/features/face-recognition/) - `insightface`-powered 1:1 verification, 1:N identification, face embedding, face detection, and demographic analysis. Ships both a non-commercial `buffalo_l` model and an Apache 2.0 OpenCV Zoo alternative.
-- **May 2026**: [Speaker diarization](/features/audio-diarization/) - new `/v1/audio/diarization` endpoint returning "who spoke when" segments. Backed by `sherpa-onnx` (pyannote-3.0 + speaker embeddings + clustering) for pure diarization, and `vibevoice-cpp` for diarization bundled with long-form ASR. Supports `json` / `verbose_json` / `rttm` response formats.
-- **June 2026**: [Sound classification](/features/audio-classification/) - new `/v1/audio/classification` endpoint for audio tagging / sound-event classification, returning scored [AudioSet](https://research.google.com/audioset/) labels (baby cry, glass breaking, alarms, ...). Backed by [ced.cpp](https://github.com/localai-org/ced.cpp), a 527-class AudioSet tagger ported to ggml.
-- **June 2026**: [PII analyze / redact API](/features/middleware/#analyze--redact-api) - the PII detection pipeline (NER + restricted-regex pattern tiers) is now a standalone service: `POST /api/pii/analyze` returns detected entity spans and `POST /api/pii/redact` returns the sanitised text (or `400 pii_blocked`), without routing a chat request through the middleware. Events gain an `origin` (`middleware` / `proxy` / `pii_analyze` / `pii_redact`) so `/api/pii/events` can be filtered by source.
-- **July 2026**: [Model capabilities endpoint](/features/api-discovery/#model-capabilities) - `GET /v1/models/capabilities`, an additive superset of `/v1/models` that reports each model's `capabilities` plus its `input_modalities` / `output_modalities` (`text` / `image` / `audio` / `video`). Lets clients route attachments using inferred or explicitly declared model modalities instead of backend-name checks.
-- **June 2026**: Concurrent scoring and PII NER on llama.cpp - the `Score` (router classifier) and `TokenClassify` (PII NER) primitives now ride llama.cpp's server task queue instead of locking the context, so they run concurrently with chat/completion/embedding traffic and with each other. The `known_usecases` restriction that forced dedicated scorer/NER model configs on llama-cpp is lifted, repeated scoring calls reuse the prompt KV cache across candidates, and scoring inputs are no longer capped by the physical batch size.
-
-## 2024 Highlights
-
-- **April 2024**: [Reranker API](https://github.com/mudler/LocalAI/pull/2121)
-- **May 2024**: [Distributed inferencing](https://github.com/mudler/LocalAI/pull/2324), [Decentralized P2P llama.cpp](https://github.com/mudler/LocalAI/pull/2343) - [Docs](https://localai.io/features/distribute/)
-- **July/August 2024**: [P2P Dashboard, Federated mode and AI Swarms](https://github.com/mudler/LocalAI/pull/2723), [P2P Global community pools](https://github.com/mudler/LocalAI/issues/3113), FLUX-1 support, [P2P Explorer](https://explorer.localai.io)
-- **October 2024**: Examples moved to [LocalAI-examples](https://github.com/mudler/LocalAI-examples)
-- **November 2024**: [Voice Activity Detection (VAD)](https://github.com/mudler/LocalAI/pull/4204), [Bark.cpp backend](https://github.com/mudler/LocalAI/pull/4287)
-- **December 2024**: [stablediffusion.cpp backend (ggml)](https://github.com/mudler/LocalAI/pull/4289)
+This page used to carry a hand-maintained highlights list. It drifted against both sources above, so it now points at them instead.

gallery/index.yaml

@@ -189,7 +189,7 @@
   files:
     - filename: DeepSeek-V4-Flash-0731-MXFP4.gguf
       uri: huggingface://ggml-org/DeepSeek-V4-Flash-0731-GGUF/DeepSeek-V4-Flash-0731-MXFP4.gguf
-      sha256: c8b46876c3939a6e141f9e4d4aa422981df4a9b84f19e9bb4e1c9a28be31e484
+      sha256: 65f73494afaf27d3add0751a5b716dd2d3e012c66ae0dbbcc1bf8477f92b3ab7
 - name: instella-moe-16b-a3b-think
   url: github:mudler/LocalAI/gallery/virtual.yaml@master
   urls:
@@ -311,7 +311,7 @@
   files:
     - filename: llama-cpp/models/Parable-Granite-4.1-3B-Claude-Fable-5-Q4_K_M/Parable-Granite-4.1-3B-Claude-Fable-5-GGUF-Q4_K_M.gguf
       uri: https://huggingface.co/AnkitAI/Parable-Granite-4.1-3B-Claude-Fable-5-GGUF/resolve/main/Parable-Granite-4.1-3B-Claude-Fable-5-GGUF-Q4_K_M.gguf
-      sha256: 67dc7695d92939c713165761f115c9d892fdff74fcbd987c8bb453b9b8ab645d
+      sha256: dbf202638af23e72508d8316577655d24ba2037fda51ce802b8996977e290bce
 - name: "parable-qwen3-4b-claude-fable-5"
   url: "github:mudler/LocalAI/gallery/virtual.yaml@master"
   urls:
@@ -345,7 +345,7 @@
   files:
     - filename: llama-cpp/models/Parable-Qwen3-4B-Claude-Fable-5-Q4_K_M/Parable-Qwen3-4B-Claude-Fable-5-GGUF-Q4_K_M.gguf
       uri: https://huggingface.co/AnkitAI/Parable-Qwen3-4B-Claude-Fable-5-GGUF/resolve/main/Parable-Qwen3-4B-Claude-Fable-5-GGUF-Q4_K_M.gguf
-      sha256: c94b06a912aa901f3da5689754577ad534415efafc50dcee3f389594a153bf38
+      sha256: 65cc4824fb78ecaf55afdfcdb6dd2e27e1aa805d289db89eae94d32d450403f0
 - name: "parable-granite-4.1-8b-claude-fable-5"
   url: "github:mudler/LocalAI/gallery/virtual.yaml@master"
   urls:
@@ -381,7 +381,7 @@
   files:
     - filename: llama-cpp/models/Parable-Granite-4.1-8B-Claude-Fable-5-Q4_K_M/Parable-Granite-4.1-8B-Claude-Fable-5-GGUF-Q4_K_M.gguf
       uri: https://huggingface.co/AnkitAI/Parable-Granite-4.1-8B-Claude-Fable-5-GGUF/resolve/main/Parable-Granite-4.1-8B-Claude-Fable-5-GGUF-Q4_K_M.gguf
-      sha256: 61a8133c344a0d0a00188395afe33c803e3b973cb4bbfd5ef1fa7110e80bc1c3
+      sha256: 57e464ae3d35253d4351639757dc35e71bab8324d12d49a5870695ce73dc19cf
 - name: "parable-qwen3-8b-claude-fable-5"
   url: "github:mudler/LocalAI/gallery/virtual.yaml@master"
   urls:
@@ -415,7 +415,7 @@
   files:
     - filename: llama-cpp/models/Parable-Qwen3-8B-Claude-Fable-5-Q4_K_M/Parable-Qwen3-8B-Claude-Fable-5-GGUF-Q4_K_M.gguf
       uri: https://huggingface.co/AnkitAI/Parable-Qwen3-8B-Claude-Fable-5-GGUF/resolve/main/Parable-Qwen3-8B-Claude-Fable-5-GGUF-Q4_K_M.gguf
-      sha256: 956070afc8023b8665fe450842f7be76b505b53d142460fd9b588222f4e16112
+      sha256: 4532d2379d38a37279866a030e51d419561f9d4d22fee00d2a33647d66f05065
 - &pocket-35b
   name: "pocket-35b"
   variants:
@@ -2009,7 +2009,7 @@
   files:
     - filename: ds4flash.gguf
       uri: https://huggingface.co/unsloth/DeepSeek-V4-Flash-GGUF
-      sha256: 856c407993ccffa9ad52e23fbef8bb7b458c792a52278f4ca7931741b0c20ce2
+      sha256: 1bfdafd1c288eb1b2bcb629ee9e1b7567dcf0abbe4d20995905a3c3465e9bd1e
 - name: "qwopus3.6-35b-a3b-coder-mtp"
   url: "github:mudler/LocalAI/gallery/virtual.yaml@master"
   urls:

website/content/blog/_index.md

@@ -1,5 +1,5 @@
 ---
 title: "Blog"
-description: "Release write-ups, benchmark reports and engineering notes from the LocalAI team. Every number here comes out of a benchmark suite, a release or a commit, and the source is named so you can check it."
+description: "Release write-ups, benchmark reports and engineering notes from the LocalAI team. Numbers link to the release, commit or benchmark run they came from."
 extracss: ["blog.css"]
 ---

website/content/blog/apex-moe-quantization.md

@@ -4,13 +4,13 @@ date: 2026-04-10
 author: "Ettore Di Giacinto"
 category: "Research"
 tags: ["quantization", "APEX", "mixture-of-experts", "llama.cpp", "benchmarks"]
-summary: "Qwen3.5-35B-A3B goes from 64.6 GB to 12.2 GB and speeds up from 30.4 to 74.4 tokens per second. Perplexity moves from 6.537 to 7.088. Here is the precision assignment that does it, and where it costs you."
+summary: "Qwen3.5-35B-A3B goes from 64.6 GB to 12.2 GB and speeds up from 30.4 to 74.4 tokens per second. Perplexity moves from 6.537 to 7.088. Here is the precision assignment that does it, and where the quality drops."
 extracss: ["blog.css"]
 ---
 
 A 35B mixture-of-experts model at full precision is a 64.6 GB file, which puts it out of reach of every consumer GPU. APEX gets Qwen3.5-35B-A3B down to 12.2 GB, where it fits a 16 GB card with room for context, and it generates at 74.4 tokens per second instead of 30.4. The output is an ordinary GGUF that stock llama.cpp opens with no patches and no custom build.
 
-The compression is not free at that tier, and the numbers below say exactly what it costs. At the 21.3 GB tier it is closer to free than we expected: APEX Quality has a lower perplexity than the F16 model it was quantized from.
+At that tier the quality does drop, and the numbers below say by how much. At the 21.3 GB tier it barely drops at all: APEX Quality has a lower perplexity than the F16 model it was quantized from.
 
 ## The measurements
 
@@ -37,17 +37,17 @@ All of this is Qwen3.5-35B-A3B on an NVIDIA DGX Spark (GB10, 122 GB unified VRAM
 
 Three things in that table are worth stopping on.
 
-APEX Quality is 21.3 GB, a third of F16, and its perplexity of 6.527 is lower than F16's 6.537 and lower than Q8_0's 6.533. Quantization noise acting as mild regularization on a wikitext evaluation is a known effect and we are not claiming the quantized model is smarter. The honest reading is that at this tier the loss is below the measurement floor.
+APEX Quality is 21.3 GB, a third of F16, and its perplexity of 6.527 is lower than F16's 6.537 and lower than Q8_0's 6.533. Quantization noise acting as mild regularization on a wikitext evaluation is a known effect and we are not claiming the quantized model is smarter. At this tier the loss is below the measurement floor.
 
-Against Unsloth's UD-Q8_K_XL, APEX I-Quality is half the size (21.3 GB against 45.3 GB), one point ahead on HellaSwag (83.5% against 82.5%), within 0.016 on perplexity, and 73% faster (63.1 t/s against 36.4). That is the comparison that matters for anyone choosing a published quant today.
+Against Unsloth's UD-Q8_K_XL, APEX I-Quality is half the size (21.3 GB against 45.3 GB), one point ahead on HellaSwag (83.5% against 82.5%), within 0.016 on perplexity, and 73% faster (63.1 t/s against 36.4).
 
 At the bottom end, APEX Mini beats bartowski IQ2_M on every metric while being 0.9 GB larger: perplexity 7.088 against 7.303, HellaSwag 81.0% against 80.3%, MMLU 41.3% against 39.6%.
 
-## Why it gets faster, not just smaller
+## Why it also gets faster
 
 Token generation on a single stream is bound by memory bandwidth, not by arithmetic. Every generated token requires reading the active weights out of memory, so halving the bytes roughly halves the time spent waiting for them. Going from 64.6 GB to 12.2 GB takes throughput from 30.4 to 74.4 tokens per second, a 2.45x gain on the same hardware with the same kernels. Every APEX tier clears 60 t/s.
 
-That is also why a large well-behaved quant such as UD-Q8_K_XL is slower than a smaller one with equal quality. Size is a speed knob as much as a memory knob.
+That is also why a large well-behaved quant such as UD-Q8_K_XL is slower than a smaller one with equal quality.
 
 ## Per-tensor and per-layer precision
 
@@ -55,9 +55,9 @@ Uniform quantization gives every tensor the same bit width, which spends the sam
 
 APEX classifies every tensor into one of three roles and treats them differently.
 
-**Routed expert weights** (the gate, up and down projections inside the experts) are the bulk of the parameters, and only 8 of 256 experts are active per token. That 97% structural sparsity is what makes aggressive quantization safe here. The routing decision itself reads full-precision gate weights, so quantization noise inside an expert that was not selected never reaches the output at all. When an expert is selected, its contribution is one of eight summed paths, which further dilutes per-tensor error.
+**Routed expert weights** (the gate, up and down projections inside the experts) are the bulk of the parameters, and only 8 of 256 experts are active per token. That 97% structural sparsity is why aggressive quantization is safe here. The routing decision itself reads full-precision gate weights, so quantization noise inside an expert that was not selected never reaches the output at all. When an expert is selected, its contribution is one of eight summed paths, which further dilutes per-tensor error.
 
-**Shared expert weights** run for every single token and their weight distribution is heavy-tailed, with a kurtosis of 13.10 against 3.41 for routed experts. Those outliers carry real signal and low-bit formats clip them. Q8_0 is the minimum viable precision here, and dropping it is the fastest way to wreck a build.
+**Shared expert weights** run for every single token and their weight distribution is heavy-tailed, with a kurtosis of 13.10 against 3.41 for routed experts. Those outliers carry real signal and low-bit formats clip them. Q8_0 is the minimum viable precision here, and dropping it degrades the build quickly.
 
 **Attention and SSM weights** are dense, contribute few parameters relative to the experts, and matter for generation quality. They sit at Q6_K throughout.
 
@@ -69,23 +69,23 @@ None of this needs a patched llama.cpp. The assignments are expressed with the s
 
 Twenty-five or so systematic runs produced a few results that saved a lot of time later.
 
-Going from Q6_K to Q8_0 on routed experts costs 7.5 GB and buys zero perplexity improvement. Going below Q5_K on them causes measurable degradation. Q6_K is the ceiling worth paying for.
+Going from Q6_K to Q8_0 on routed experts costs 7.5 GB and gives zero perplexity improvement. Going below Q5_K on them causes measurable degradation. Q6_K is the ceiling.
 
 Layer position matters more than uniform bit width. A two-tier gradient of Q6_K edges and Q5_K middle matches Q8_0 quality; a uniform Q5_K assignment at a similar size does not.
 
 IQ formats underperform K-quants on MoE experts. IQ3_S gives worse perplexity than Q3_K on routed expert tensors at a similar bit rate, because the near-Gaussian expert weight distribution (kurtosis 3.41) suits the K-quant block structure better.
 
-Five C-level modifications to the quantization algorithms themselves, including error feedback, enhanced scale search, super-block refinement and Gaussian-density weighting, all showed zero improvement. Stock llama.cpp quantization is already good. The gains here come entirely from deciding where to spend bits.
+Five C-level modifications to the quantization algorithms themselves, including error feedback, enhanced scale search, super-block refinement and Gaussian-density weighting, all showed zero improvement. Stock llama.cpp quantization is already good. The gains here come entirely from deciding where to put the bits.
 
 ## The I-variants and their calibration set
 
 Standard imatrix calibration uses Wikipedia text, which is also what wikitext perplexity measures, so the calibration and the benchmark agree with each other by construction. The I-variants calibrate on a diverse set spanning chat, code, reasoning and tool-calling, with no Wikipedia in it.
 
-That trade shows up clearly. I-Compact drops perplexity from 6.783 to 6.669, cuts KL max from 7.56 to 5.50, and lifts MMLU from 40.9% to 41.7%. At the Quality tier, I-Quality gives up 0.025 perplexity against Quality and takes the highest HellaSwag score of anything tested (83.5%), the best TruthfulQA (38.4%), and a lower KL divergence. If your workload is chat, code or agents rather than encyclopedic prose, take the I variant.
+It shows up in the numbers. I-Compact drops perplexity from 6.783 to 6.669, cuts KL max from 7.56 to 5.50, and lifts MMLU from 40.9% to 41.7%. At the Quality tier, I-Quality gives up 0.025 perplexity against Quality and takes the highest HellaSwag score of anything tested (83.5%), the best TruthfulQA (38.4%), and a lower KL divergence. If your workload is chat, code or agents rather than encyclopedic prose, take the I variant.
 
-## Where it costs you
+## Where the quality drops
 
-The Compact and Mini tiers are real compression, and they are not free.
+The Compact and Mini tiers lose real quality.
 
 Compact at 16.1 GB moves perplexity from 6.537 to 6.783, a 3.8% increase, and its KL mean rises tenfold against Q8_0, from 0.0046 to 0.0469. Mini at 12.2 GB goes to 7.088, an 8.4% increase, with a KL mean of 0.0870 and HellaSwag down 1.5 points to 81.0%. Those are the numbers to weigh against the fact that the model now runs at all on a 16 GB card.
 

website/content/blog/localai-since-march-2023.md

@@ -4,7 +4,7 @@ date: 2026-07-29
 author: "Ettore Di Giacinto"
 category: "History"
 tags: ["history", "architecture", "releases", "community"]
-summary: "Three years, 133 releases and 224 contributors later. The four changes that mattered most were making the core small, adding agents, making it a cluster, and giving it eyes and ears."
+summary: "Three years, 133 releases and 224 contributors later. Here are the four decisions that shaped it: making the core small, adding agents, making it a cluster, and giving it eyes and ears."
 extracss: ["blog.css"]
 ---
 
@@ -16,9 +16,9 @@ None of those numbers are rounded up. You can read every one of them off the rep
 
 {{< starchart >}}
 
-The curve is not the point, but it is a useful map. The four marks on it are the four decisions below, and you can see each of them in the slope afterwards.
+The four marks on it are the four decisions below, and you can see each of them in the slope afterwards.
 
-What follows is how it got here. Not the feature list, which you can read in the releases, but the four decisions that changed the shape of the thing.
+What follows is the four decisions that changed the shape of the thing. The full feature list is in the releases.
 
 ## 2023 to 2024: an API in front of llama.cpp
 
@@ -34,7 +34,7 @@ Every backend moved out of the main binary in [v3.2.0](https://github.com/mudler
 
 You install one thing and it stays small. Ask for a GGUF model and llama-cpp arrives. Ask for transcription and whisper or parakeet arrives. Nothing else is fetched, and a machine that only ever serves one model never downloads the other sixty-nine backends.
 
-That one change is what made everything after it possible. Adding a backend stopped meaning adding weight to everybody's install, so "should we support this engine" stopped being an argument about download size and went back to being an argument about whether the engine is any good. It is also the reason we can afford to maintain eighteen engines of our own, which comes later.
+Everything after it depended on that one change. Adding a backend stopped meaning adding weight to everybody's install, so "should we support this engine" stopped being an argument about download size and went back to being an argument about whether the engine is any good. It is also the reason we can afford to maintain eighteen engines of our own, which comes later.
 
 ## March 2026: agents, and a new interface
 
@@ -42,7 +42,7 @@ That one change is what made everything after it possible. Adding a backend stop
 
 The web interface was rewritten in React at the same time, with a Canvas mode, MCP Apps and client-side tools with tool streaming ([#8947](https://github.com/mudler/LocalAI/pull/8947)), and WebRTC realtime audio ([#8790](https://github.com/mudler/LocalAI/pull/8790)). MLX gained a distributed mode ([#8801](https://github.com/mudler/LocalAI/pull/8801)).
 
-The realtime audio path is the piece that changed what people built. Speech in, tool calls in the middle, speech out, over WebRTC, fast enough that it feels like a conversation rather than a walkie-talkie. It had landed as the Realtime API in February 2026 ([#6245](https://github.com/mudler/LocalAI/pull/6245)), and the interface rewrite finally gave it a face.
+The realtime audio path changed what people built with it. Speech in, tool calls in the middle, speech out, over WebRTC, fast enough that it feels like a conversation rather than a walkie-talkie. It had landed as the Realtime API in February 2026 ([#6245](https://github.com/mudler/LocalAI/pull/6245)), and the interface rewrite finally gave it a face.
 
 ## April 2026: it becomes a cluster
 
@@ -74,6 +74,6 @@ The most recent one is [vllm.cpp](https://github.com/mudler/vllm.cpp), a C++20 p
 
 ## Where it stands
 
-Still MIT, still a community project. 224 people have put code in, and the README is kept translated into eight languages because the people using this are not all in one place. The [contributors graph](https://github.com/mudler/LocalAI/graphs/contributors) is the honest picture of who actually built this, and it is not me.
+Still MIT, still a community project. 224 people have put code in, and the README is kept translated into eight languages because the people using this are not all in one place. The [contributors graph](https://github.com/mudler/LocalAI/graphs/contributors) shows who actually built this, and it is not me.
 
 If you want to add something, backends and gallery entries are the two places a first contribution lands cleanly. There is a step-by-step checklist for a new backend in `.agents/adding-backends.md`, and a gallery entry is just a YAML block. Come say hello in [Discord](https://discord.gg/uJAeKSAGDy) if you get stuck.

website/content/blog/parakeet-cpp-asr-on-cpu.md

@@ -1,22 +1,22 @@
 ---
-title: "parakeet.cpp: NeMo transcripts, byte for byte, without the Python"
+title: "parakeet.cpp: the same NeMo transcript, without the Python"
 date: 2026-06-05
 author: "Ettore Di Giacinto"
 category: "Benchmarks"
 tags: ["parakeet.cpp", "ASR", "ggml", "streaming", "benchmarks"]
-summary: "Same transcript as NVIDIA NeMo, character for character, at a median 1.40x on CPU and about 27x the speed of whisper.cpp. One binary, one GGUF file, no Python at inference."
+summary: "The same transcript as NVIDIA NeMo at a median 1.40x on CPU, and about 27x the speed of whisper.cpp, from one binary and one GGUF file."
 extracss: ["blog.css"]
 ---
 
-You can drop a single binary and a GGUF file onto a machine with no GPU and get NVIDIA NeMo Parakeet transcription out of it, at a median 1.40x NeMo's own PyTorch CPU speed, with a transcript that matches NeMo character for character. That is [parakeet.cpp](https://github.com/mudler/parakeet.cpp), a C++17 port of the Parakeet speech-recognition family built on ggml.
+You can drop a single binary and a GGUF file onto a machine with no GPU and get NVIDIA NeMo Parakeet transcription out of it, at a median 1.40x NeMo's own PyTorch CPU speed, with the same transcript NeMo produces. That is [parakeet.cpp](https://github.com/mudler/parakeet.cpp), a C++17 port of the Parakeet speech-recognition family built on ggml.
 
-Accuracy came first and speed came second, in that order, because a faster transcriber that disagrees with the reference is a different model, not a port.
+We checked the accuracy before touching the speed, because a transcriber that disagrees with the reference is not a port of it.
 
 ## WER 0 against NeMo
 
 Every published checkpoint is validated at WER 0 against NeMo. Across the LibriSpeech test-clean set the mean f32 agreement WER, meaning the word error rate between our transcript and NeMo's on the same audio, is 0.0155%. On seven of the ten models it is exactly 0.0000%, which is a byte-identical transcript.
 
-That number is what makes the speed comparison meaningful. Both engines did the same work and produced the same output, so the only difference left is how long they took.
+Both engines did the same work and produced the same output, so the only difference left is how long they took.
 
 ## CPU, against NeMo's own runtime
 
@@ -54,25 +54,25 @@ Against whisper.cpp turbo on the same clip and at the same accuracy (1.6% WER on
 
 The decisive win was on the decode side. A transducer decodes autoregressively, and profiling showed the prediction-network LSTM taking about 97% of RNN-T decode time while producing the same output over and over: on a non-emitting frame the prediction network's input has not changed, so its forward pass is redundant. Caching that forward across non-emitting frames removed most of the decode cost.
 
-The encoder side is a set of smaller wins with no single hero: a persistent ggml backend with `gallocr`, zero-copy weights straight out of the GGUF mapping, one fused graph rather than per-layer graph building, and tinyBLAS through `GGML_LLAMAFILE`.
+The encoder side is a set of smaller wins: a persistent ggml backend with `gallocr`, zero-copy weights straight out of the GGUF mapping, one fused graph rather than per-layer graph building, and tinyBLAS through `GGML_LLAMAFILE`.
 
 ## On the GPU
 
 On an NVIDIA GB10 (Grace-Blackwell), parakeet.cpp wins on all ten models, with a median of 1.25x and up to 4.3x on the large TDT and hybrid models. The reference here is NeMo-GPU inside the `nvcr.io/nvidia/nemo` container, because NeMo cannot run on that host's torch and CUDA stack directly.
 
-The 4.3x cases have a specific cause. NeMo's TDT greedy decode is not CUDA-graph accelerated and falls back to a per-step Python loop, while ours is a lean C++ loop. Where NeMo's decode is CUDA-graph accelerated, as it is for RNN-T, the gap narrows to about 1.16x at f32 and 1.30x at q8_0. On the pure-encoder CTC models the margin is around 1.2x, because ggml's generic CUDA conv and attention kernels still trail NVIDIA's tuned cuDNN. That is the main piece of GPU headroom left in the project and we say so in the README rather than averaging it away.
+The 4.3x cases have a specific cause. NeMo's TDT greedy decode is not CUDA-graph accelerated and falls back to a per-step Python loop, while ours is a lean C++ loop. Where NeMo's decode is CUDA-graph accelerated, as it is for RNN-T, the gap narrows to about 1.16x at f32 and 1.30x at q8_0. On the pure-encoder CTC models the margin is around 1.2x, because ggml's generic CUDA conv and attention kernels still trail NVIDIA's tuned cuDNN. That is the main piece of GPU headroom left in the project, and the README lists it per model.
 
-Batching several clips through the decoder together reaches about 10x to 12x at batch size 16 on the GB10, and about 3x to 5x on CPU. It applies to transducer models only, since CTC has no autoregressive decode to batch, and the batched path is bit-identical to running the clips one at a time.
+Batching several clips through the decoder together reaches about 10x to 12x at batch size 16 on the GB10, and about 3x to 5x on CPU. It applies to transducer models only, since CTC has no autoregressive decode to batch, and the batched path produces the same output as running the clips one at a time.
 
 On Apple M4 through ggml's Metal backend, the larger models run about 3x to 5x faster than the same models on that machine's CPU.
 
-## Cache-aware streaming, and what end-of-utterance detection buys you
+## Cache-aware streaming and end-of-utterance detection
 
 Offline transcription hands you a file and waits. A voice assistant cannot do that, so `parakeet_realtime_eou_120m-v1` runs a cache-aware streaming path instead: you feed it 16 kHz mono PCM as it arrives and it returns newly finalized text as it becomes stable.
 
-Cache-aware means the cost per chunk stays flat. Each chunk's forward pass carries per-layer convolution and attention caches plus the transducer decoder state forward, so nothing before the current chunk is recomputed. Without that, every chunk would re-run the encoder over the whole session so far, and the per-chunk cost would grow with the length of the conversation until the loop fell behind. The implementation covers layer norm with causal convolution, causal subsampling, and chunked-limited attention, and its transcript matches NeMo's own cache-aware streaming byte for byte.
+Cache-aware means the cost per chunk stays flat. Each chunk's forward pass carries per-layer convolution and attention caches plus the transducer decoder state forward, so nothing before the current chunk is recomputed. Without that, every chunk would re-run the encoder over the whole session so far, and the per-chunk cost would grow with the length of the conversation until the loop fell behind. The implementation covers layer norm with causal convolution, causal subsampling, and chunked-limited attention, and its transcript matches NeMo's own cache-aware streaming exactly.
 
-End-of-utterance detection is the part that changes how an assistant feels. The model emits `<EOU>` when the speaker has finished a turn and `<EOB>` for a backchannel, as events alongside the text. A voice loop can start generating a reply the moment `<EOU>` arrives rather than waiting out a fixed silence timer, which is where most of the perceived lag in a spoken assistant comes from. The alternative, a VAD with a 700 ms hangover, either cuts people off mid-sentence or makes the assistant feel slow, and it cannot tell "mm-hm" from the end of a thought. `finalize` flushes the tail at end of stream without fabricating an `<EOU>` that NeMo would not have emitted.
+End-of-utterance detection changes how an assistant feels. The model emits `<EOU>` when the speaker has finished a turn and `<EOB>` for a backchannel, as events alongside the text. A voice loop can start generating a reply the moment `<EOU>` arrives rather than waiting out a fixed silence timer, which is where most of the perceived lag in a spoken assistant comes from. The alternative, a VAD with a 700 ms hangover, either cuts people off mid-sentence or makes the assistant feel slow, and it cannot tell "mm-hm" from the end of a thought. `finalize` flushes the tail at end of stream without fabricating an `<EOU>` that NeMo would not have emitted.
 
 The streaming path measures at RTFx 3.80 on a 7.43 second clip. That sits well below the offline number by design, because streaming runs many small chunked passes rather than one large one, and it is still several times faster than real time on a CPU.
 
@@ -95,7 +95,7 @@ parakeet.cpp ports NeMo's `rel_pos_local_attn`, a banded attention where each qu
 </table>
 </div>
 
-At NeMo's full W=128 window that is about 4x faster and about 5.7x less peak memory than the global path. The band is built with a chunk-matmul construction, overlapping key and value chunks feeding one batched GEMM plus a diagonal skew view, so the graph node count does not depend on the window. The wide window costs the same as the narrow one. Short clips stay on the global path and remain byte-identical to before.
+At NeMo's full W=128 window that is about 4x faster and about 5.7x less peak memory than the global path. The band is built with a chunk-matmul construction, overlapping key and value chunks feeding one batched GEMM plus a diagonal skew view, so the graph node count does not depend on the window. The wide window costs the same as the narrow one. Short clips stay on the global path and produce the same output as before.
 
 ## Using it
 

website/content/blog/what-landed-in-localai-3-10.md

@@ -0,0 +1,59 @@
+---
+title: "LocalAI 3.10: the Anthropic and Responses APIs, and one image for every GPU"
+date: 2026-01-18
+author: "Ettore Di Giacinto"
+category: "Release"
+tags: ["release", "anthropic", "open-responses", "gpu", "moonshine"]
+summary: "A /v1/messages endpoint that Claude clients can talk to unchanged, Open Responses compatibility that passes the official acceptance tests, and GPU libraries moved inside the backend containers so one image works on any hardware."
+extracss: ["blog.css"]
+---
+
+Half the tooling worth using speaks a shape of API that is not OpenAI's. You find a client you like, it talks to Anthropic, and swapping it onto a local model means either rewriting the client or gluing a translation layer in front of it. Same story with the agent frameworks that went all in on the Responses API.
+
+3.10.0 adds both surfaces natively, so the client does not have to know.
+
+## Two more front doors
+
+The Anthropic Messages API is served at `/v1/messages`, and at `/messages` for clients that do not prefix. Tool calling, streaming and non-streaming all work, so `anthropic-sdk-go`, LangChain and anything else built on that shape can be pointed at your instance without a code change.
+
+The Open Responses API is at `/v1/responses`, with `/v1/responses/:id` to fetch one and `/v1/responses/:id/cancel` to stop it. It is stateful: pass a `response_id` and the conversation resumes, set `background: true` and the agent runs asynchronously while you go and do something else, then come back for the result. Streaming covers tools, images and audio.
+
+That one passes the [official acceptance tests](https://www.openresponses.org/compliance), which was the bar I wanted to hit before shipping it.
+
+## One image for every GPU
+
+This is the change most likely to affect you even if you do not care about agents.
+
+GPU libraries (CUDA, ROCm, Vulkan) now live inside the backend containers rather than in the image you pull. There is no longer a CUDA image, a ROCm image and a CPU image to choose between. You pull the image, and acceleration works if the hardware is there! Vulkan arm64 builds are in too.
+
+It is experimental, and I want to be clear about that rather than bury it. It is a real architectural change to how every backend gets its libraries, and there will be hardware combinations we did not hit. If it does not work on yours, please file an issue, that is genuinely the most useful thing you can do for this one.
+
+## Everything else
+
+The backend gallery is system aware now, so it only lists backends your machine can actually run. No more scrolling past MLX entries on a Linux box.
+
+Tool calls stream properly, including partial arguments as `input_json_delta`, and models that emit tools as XML (`<function>...</function>`) get parsed instead of dumping the markup into the message text. Both work across llama.cpp, vLLM and diffusers.
+
+Thinking tags are extracted into a separate `reasoning` field rather than being left in the answer, in both SSE and non-SSE mode. The chat UI shows them under a Thinking tab.
+
+There is a video generation page in the web UI with LTX-2 behind it, doing text-to-video and image-to-video with the usual `fps`, `num_frames` and `guidance_scale` controls.
+
+There is request tracing now. `GET /api/traces` returns in-memory request and response logs, `/api/traces/clear` empties them. It is memory backed and drops old entries past a size cap, so it is for debugging an agent that is misbehaving right now, not for an audit trail.
+
+Two new speech backends. Moonshine is an ONNX transcription engine aimed at low-end hardware, and it is the one to reach for on a Pi or an old laptop. It is quick! Pocket-TTS does lightweight TTS with voice cloning, though the cloning path needs a HuggingFace login and a registered voice model, so it is not quite copy-paste.
+
+## Old hardware, and AMD memory
+
+Two fixes worth calling out because they were silent failures rather than errors.
+
+LocalAI was crashing on Intel CPUs without BMI2 (Sandy Bridge, Ivy Bridge), showing up as an `EOF` during model warmup rather than anything that pointed at the cause. It now falls back to `llama-cpp-fallback` on those chips.
+
+On AMD, used and total VRAM were swapped when parsing `rocm-smi` output, so a dual-Radeon box reported nonsense. `HIP_VISIBLE_DEVICES` is also handled properly now, which matters if you are pinning to the discrete GPU.
+
+## Thanks
+
+Thanks to @richiejp, @majiayu000, @nanoandrew4, @DEVMANISHOFFL, @coffeerunhobby, @rampa3, @Nold360, @jroeber and @Divyanshupandey007 for the work in this cycle.
+
+If the unified GPU backends misbehave on your setup, open an issue with what hardware you are on. And if you are wiring up the Anthropic or Responses endpoints and something does not match the spec, tell me, I would rather hear it from you than find out later.
+
+[Full release notes](https://github.com/mudler/LocalAI/releases/tag/v3.10.0).

website/content/blog/what-landed-in-localai-4-0.md

@@ -0,0 +1,69 @@
+---
+title: "LocalAI 4.0: agents in the core, and a React interface"
+date: 2026-03-14
+author: "Ettore Di Giacinto"
+category: "Release"
+tags: ["release", "agents", "agenthub", "mcp", "react", "webrtc"]
+summary: "Native agent orchestration with the Agenthub, a rewritten interface with Canvas mode, MCP Apps with tool streaming, and two things removed."
+extracss: ["blog.css"]
+---
+
+Running an agent locally has meant running two things: an inference server, and a separate orchestrator that talks to it. That is a lot of moving parts for something you wanted to try on a Tuesday evening.
+
+4.0.0 puts the agent side in the core. You create agents, give them memory and skills, connect them to MCP servers, and start and stop them from the same interface you already use for models.
+
+This is a major version bump, so there are two removals near the bottom of this post. Read those before you upgrade.
+
+## Agents, and the Agenthub
+
+Agents are managed through the React interface: create one, wire up MCP servers and skills, connect it to Slack, watch what it is doing through a new Events column in the agents list.
+
+Memory has two options. Hybrid search backed by PostgreSQL if you already run one, or in-memory storage via Chromem if you do not want another service. Skills live in a central database rather than being pasted per agent.
+
+The bit I am most curious to see used is [Agenthub](https://agenthub.localai.io), a community space for sharing agent configurations. You publish one, somebody else imports it into their instance and runs it against their own models on their own hardware!
+
+## The interface is React now
+
+The web interface has been rewritten. The old one had reached the point where adding anything meant fighting it.
+
+Canvas mode is the new thing worth turning on: enable it in chat and code blocks and artifacts the model produces render in a preview pane on the right instead of scrolling past you as text. The System view splits Models and Backends into tabs. Traces render as accordions, which makes a long one readable. And if you try to install a model whose weights exceed your system RAM, you get a warning first rather than a locked-up machine.
+
+## MCP Apps
+
+Client-side MCP support is complete in this release ([#8947](https://github.com/mudler/LocalAI/pull/8947)). You pick which MCP servers to enable for a chat directly in the interface, and their tools get injected into the normal chat with streaming, so there is no separate agent mode to switch into.
+
+If you would rather not have any of it, `LOCALAI_DISABLE_MCP` turns the whole thing off.
+
+## Audio, video, and MLX across machines
+
+WebRTC is wired into the Realtime API and the Talk page ([#8790](https://github.com/mudler/LocalAI/pull/8790)), which is a real improvement for latency over what was there before.
+
+Three new audio backends: fish-speech, ace-step.cpp, and faster-qwen3-tts (CUDA only). TTS gained `sample_rate` support through post-processing, and Qwen TTS handles multiple voices.
+
+There is also an experimental MLX distributed backend for spreading a workload across Apple machines ([#8801](https://github.com/mudler/LocalAI/pull/8801)). It is early, so expect rough edges if you try it.
+
+## Infrastructure
+
+Persistent data now has its own location, separate from configuration. `LOCALAI_DATA_PATH` (or `--data-path`) points at where agents, skills, tasks, jobs and the collection database live, defaulting to `data/` under the base path. If you are mounting volumes, this is the one to look at.
+
+Shell completion scripts generate for bash, zsh and fish. There is dedicated Podman documentation now, including rootless setup.
+
+## Two things are gone
+
+The HuggingFace backend has been removed.
+
+AIO images are dropped. They existed to bundle a preset of models with the runtime, and maintaining them across every hardware variant stopped being worth what they gave people. Use the main images and install models from the gallery.
+
+## One known issue
+
+The `diffusers` backend is not in this release. It failed to build because we exhausted our CI limits, so the previous version is still what you get if you install it.
+
+This is an infrastructure problem, not a code one, and it is the kind of thing that will keep happening to us. If you know anybody at GitHub who could help us get better ARM runners, please reach out, I am not too proud to ask.
+
+## Thanks
+
+Thanks to @richiejp, @nanoandrew4, @Weathercold, @sozercan, @lukasdotcom, @loryanstrant, @bittoby and @attilagyorffy.
+
+If you build an agent worth sharing, put it on the Agenthub. The more the merrier!
+
+[Full release notes](https://github.com/mudler/LocalAI/releases/tag/v4.0.0).

website/content/blog/what-landed-in-localai-4-1.md

@@ -0,0 +1,76 @@
+---
+title: "LocalAI 4.1: more than one box, and more than one user"
+date: 2026-04-02
+author: "Ettore Di Giacinto"
+category: "Release"
+tags: ["release", "distributed", "auth", "oidc", "quotas", "fine-tuning"]
+summary: "Distributed cluster mode that places requests by real free VRAM, OIDC with per-user API keys and quotas, and LoRA fine-tuning that exports straight to GGUF."
+extracss: ["blog.css"]
+---
+
+Two problems show up the moment LocalAI stops being a thing you run for yourself.
+
+The first is that you have more than one machine, and only one of them is doing any work. The second is that other people are using your instance, and you have no way to tell who is burning the GPU, or to stop them.
+
+4.1.0 is mostly about those two.
+
+## Running as a cluster
+
+Distributed mode lets you point several nodes at one control plane and stop thinking about which one to call.
+
+Routing orders nodes by available VRAM, so the request lands on the card with room for it. Node groups let you pin models to a subset of the cluster, which is how you keep a heavy diffusion model off the boxes doing embeddings. There is a min/max autoscaler with a reconciler managing node lifecycle, and you can drain a node for maintenance and resume it later through the API instead of pulling it out from under in-flight requests.
+
+Model transfer between nodes goes over S3 or peer to peer, so a model you have already pulled once does not have to come down from the internet again on every node!
+
+The cluster status shows up on the home page.
+
+## Users, keys and quotas
+
+LocalAI ships a multi-user platform now, which is the piece that makes it deployable for a team or a classroom rather than just for you.
+
+- User management from the React interface.
+- OIDC/OAuth against your own identity provider (Google, Keycloak, Authentik, whatever you already run).
+- Invite mode, so registration is closed unless an admin lets somebody in.
+- Per-user API keys.
+- Admin impersonation, for when somebody reports a bug you cannot reproduce.
+
+On top of that there is a quota system: set per-user limits and have them enforced, with a usage dashboard broken down per user and a predictive view of where consumption is heading.
+
+## Fine-tuning without leaving the interface
+
+Both of these are experimental. I would use them on something you can afford to throw away.
+
+Fine-tuning uses HuggingFace TRL to train LoRA adapters, exports the result to GGUF automatically, and imports it back into LocalAI so you can serve what you just trained without moving files around by hand. There is a small evals framework included to check whether the thing you trained is actually better.
+
+The quantization backend produces optimized variants of a model on the fly.
+
+## Agents from the terminal
+
+You can run an agent without the server now:
+
+```sh
+local-ai agent run <name>
+local-ai agent list
+```
+
+`run` takes an agent from the pool registry in `pool.json`, or a single-turn `--prompt` if you just want one answer. Tool calls stream in real time, and the interleaved-thinking bug that mangled output when a model reasoned mid-tool-call is fixed.
+
+## The rest of the interface work
+
+The model pipeline editor is visual, so wiring models together no longer means editing YAML. Backend logs can be scoped to a single model rather than reading the whole stream. Studio pages remember past generations, so images and audio you made last week are still there. The model and backend selectors are searchable. Error toasts link straight to the trace that produced them.
+
+## Under the hood
+
+Inference defaults are pulled from Unsloth and applied across all endpoints and gallery models, so models arrive with sane sampling parameters instead of whatever the default happened to be. `min_p` is supported. When native tool-call parsing fails, an iterative fallback parser takes over rather than returning nothing.
+
+Repeated log lines get collapsed. NVIDIA Jetson and Tegra are detected as first-class platforms. SYCL backends auto-disable `mmap`, which was crashing them on Intel GPUs. llama.cpp bundles `libdl`, `librt` and `libpthread` for portability. And the downloader rewrites HuggingFace URIs through `HF_ENDPOINT`, which is the one you need if you are behind a corporate mirror.
+
+## Thanks
+
+Thanks to @richiejp for a large chunk of this cycle, and to @tv42, @walcz-de, @majiayu000 and @ER-EPR.
+
+There is a full setup walkthrough on video if you would rather watch than read: [youtube.com/watch?v=cMVNnlqwfw4](https://www.youtube.com/watch?v=cMVNnlqwfw4).
+
+If you are setting up distributed mode or OIDC and hit a wall, reach out, I am happy to help you get it standing up.
+
+[Full release notes](https://github.com/mudler/LocalAI/releases/tag/v4.1.0).

website/content/blog/what-landed-in-localai-4-2.md

@@ -0,0 +1,117 @@
+---
+title: "LocalAI 4.2: who spoke when, and whose face is that"
+date: 2026-05-11
+author: "Ettore Di Giacinto"
+category: "Release"
+tags: ["release", "diarization", "voice-recognition", "face-recognition", "ollama", "backends"]
+summary: "A /v1/audio/diarization endpoint, voice and face recognition with liveness, a drop-in Ollama API, and eleven new backends."
+extracss: ["blog.css"]
+---
+
+You record an hour of standup, run it through Whisper, and get back one long wall of text. Every word is correct. You still have no idea who said any of them, so you end up scrubbing through the audio with the transcript open in another window, guessing at voices.
+
+4.2.0 is mostly about that class of problem. Audio and images carry more than "here are the words" or "here is a picture", and until now LocalAI had nowhere to put the rest of it.
+
+## Who spoke when
+
+There is a new `/v1/audio/diarization` endpoint, shaped like `/v1/audio/transcriptions` so your existing multipart code mostly carries over:
+
+```bash
+curl http://localhost:8080/v1/audio/diarization \
+  -H "Content-Type: multipart/form-data" \
+  -F file="@meeting.wav" \
+  -F model="vibevoice-cpp-asr" \
+  -F num_speakers=3
+```
+
+```json
+{
+  "task": "diarize",
+  "duration": 12.34,
+  "num_speakers": 2,
+  "segments": [
+    {"id": 0, "speaker": "SPEAKER_00", "label": "0", "start": 0.00, "end": 2.34},
+    {"id": 1, "speaker": "SPEAKER_01", "label": "1", "start": 2.34, "end": 4.10}
+  ]
+}
+```
+
+Two backends serve it. [sherpa-onnx](https://github.com/k2-fsa/sherpa-onnx) does pure diarization (pyannote-3.0 segmentation, a speaker-embedding extractor, then clustering) and never transcribes, so you do not pay for ASR you did not ask for. `vibevoice-cpp` emits speaker-labelled segments as a by-product of its long-form ASR pass, so with `include_text=true` you get a transcript per segment for free! `response_format` gives you `json`, `verbose_json`, or `rttm` if you want to feed the output to `dscore`.
+
+One thing to know before you build on it: `SPEAKER_00` is local to a single request. Run the same meeting twice and the numbering can come out differently, and nothing promises that `SPEAKER_00` in Monday's recording is the same human as `SPEAKER_00` in Tuesday's. If you need identity across files, pair it with `/v1/voice/embed` and keep your own embedding store. Which brings me to..
+
+## Voices and faces
+
+`/v1/voice/*` is new ([#9500](https://github.com/mudler/LocalAI/pull/9500)): verify (are these two clips the same person?), identify (which of my enrolled speakers is this?), embed (give me the vector, I will do the rest myself), and analyze (age, gender, emotion).
+
+```bash
+local-ai models install speechbrain-ecapa-tdnn
+
+curl -sX POST http://localhost:8080/v1/voice/verify \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "speechbrain-ecapa-tdnn",
+    "audio1": "https://example.com/alice_1.wav",
+    "audio2": "https://example.com/alice_2.wav"
+  }'
+```
+
+```json
+{"verified": true, "distance": 0.18, "threshold": 0.25}
+```
+
+The default threshold is around 0.25 for ECAPA-TDNN, and it moves per engine, so pass `threshold` explicitly if you swap the model out.
+
+`/v1/face/*` does the same thing for faces ([#9480](https://github.com/mudler/LocalAI/pull/9480)), plus detection and demographics, and 4.2.0 adds antispoofing. Holding a printed photo or a phone screen up to the camera is the oldest attack on face auth there is, and the liveness check rejects it.
+
+Some honest limits. Liveness is an arms race and this is not bank-grade. The demographic heads emit confident-looking numbers for age and emotion that you should read as a rough signal and not as a fact about a person. And the default `insightface` buffalo packs are released for non-commercial research use only, so if you are shipping this in a product, pick the OpenCV Zoo entry instead. That is in the docs, but people skip docs, so it is here too.
+
+The samples never leave your machine, which is the part I actually care about. They go from your process to the backend running next to it and nowhere else. Doing biometrics against somebody else's cloud API always felt like the worst possible trade.
+
+## Point your ollama client at LocalAI
+
+```sh
+OLLAMA_HOST=http://localhost:8080 ollama run qwen3
+```
+
+LocalAI answers the Ollama API now ([#9284](https://github.com/mudler/LocalAI/pull/9284)), so a tool that only ever learned to talk to Ollama keeps working with no code change on your side. `/api/chat`, `/api/generate`, `/api/embed`, `/api/tags`, `/api/show`, `/api/ps` and `/api/version` all land on the engine you were already running, and your existing `/v1/*` clients are untouched.
+
+There is no `/api/pull` in there. Models come from the LocalAI gallery or from a URL you hand it, so `ollama run` against something you have not installed yet will not go and fetch it for you.
+
+## Video, and an interface repaint
+
+`stable-diffusion.ggml` generates video now ([#9420](https://github.com/mudler/LocalAI/pull/9420))! There are gallery entries for Wan 2.1 FLF2V 14B 720P and Wan i2v 720p, including first-last-frame interpolation.
+
+The React interface got a long cycle of work. The chat is redesigned, the palette moved to Nord, and there is i18n across English, Italiano, Español, Deutsch and 简体中文. You can brand your instance too - name, tagline, logo, favicon - and the login page, sidebar, footer and browser tab all pick it up. Handy if you run LocalAI for a team and would rather it did not look like somebody's side project.
+
+The model config editor is interactive now, with autocomplete over known fields and live validation, and it renames the file on save so you stop accumulating three copies of the same config.
+
+## Eleven new backends
+
+sglang, ik-llama.cpp, TurboQuant, sam.cpp, Kokoros, qwen3tts.cpp, tinygrad-multimodal (experimental, do not build anything load-bearing on it yet), vibevoice.cpp, LocalVQE, insightface, and voice-rec.
+
+vLLM reached feature parity with llama.cpp in this cycle. The full `AsyncEngineArgs` surface is exposed as a generic YAML map, and tensor-parallel distributed workers let a single model span nodes. There are CUDA 13 builds for vLLM, vLLM-omni and sglang, plus L4T arm64 for Jetson-class boards.
+
+## The unglamorous half
+
+Most of the 279 pull requests here are not features. A sample of what actually went in:
+
+- llama.cpp renamed its `common` target to `llama-common`, which broke the TurboQuant build until the detection was fixed.
+- ik-llama.cpp needed a patch to `clip.cpp` for the new `ggml_quantize_chunk` signature, plus adapting to the `common_grammar` struct in `sampling.h`.
+- `mlx-vlm` is pinned to v0.4.4 to unblock CUDA builds.
+- vLLM dropped the flash-attn wheel to avoid a torch 2.10 ABI mismatch.
+- Whisper transcriptions can be cancelled by the client, through the ggml `abort_callback`, so aborting a request frees the GPU instead of letting it run to completion in the background.
+- faster-whisper emits word-level timestamps.
+- gfx1151 (Strix Halo / Ryzen AI MAX) works, with `AMDGPU_TARGETS` exposed as a build-arg.
+
+On the security side: an unsafe `sprintf()` came out of the C++ grpc-server, env-supplied API keys are stripped from Settings API requests before they get persisted so they cannot leak back out through the config, and deleting a user on PostgreSQL cascades across everything they owned instead of leaving orphaned rows behind.
+
+Distributed mode got a hardening pass. Round-robin across replicas of the same model, "Upgrade All" scoped to the nodes that actually have the backend installed, NATS `backend.upgrade` split off from install, and correct VRAM/RAM reporting on NVIDIA unified-memory hosts.
+
+## Thanks
+
+This one had a lot of hands on it. Thanks to @richiejp for the model config editor, Kokoros and a pile of build fixes, @Anai-Guo, @russell, @leinasi2014, @keithmattix for gfx1151, @orbisai0security and @SAY-5 for the security work, @walcz-de, @thelittlefireman, @sec171, @pjbrzozowski, @mvanhorn, @arteven, @Dennisadira, @eglia, @arbrick, @neurocis and @ER-EPR.
+
+If you are wiring up diarization or the voice endpoints and get stuck, open an issue or reach out, I am genuinely happy to help you get it working.
+
+[Full release notes](https://github.com/mudler/LocalAI/releases/tag/v4.2.0).

website/content/blog/what-landed-in-localai-4-3.md

@@ -0,0 +1,105 @@
+---
+title: "LocalAI 4.3: signed backends, and the prompt cache that was off"
+date: 2026-05-24
+author: "Ettore Di Giacinto"
+category: "Release"
+tags: ["release", "security", "cosign", "prompt-cache", "distributed", "usage"]
+summary: "Keyless cosign verification for backend OCI images, the llama.cpp prompt cache enabled by default, per-API-key usage attribution, and the replica-pinning bug that kept a second node idle."
+extracss: ["blog.css"]
+---
+
+Here is a gap that had been sitting in LocalAI for a while. The gallery YAML tells LocalAI which OCI image to pull for a backend, and then LocalAI pulls it. Nothing checked that the bytes coming back were the bytes we built. A compromised registry, or somebody in the middle, and you would never know.
+
+4.3.0 closes that, and fixes a default that had been quietly costing everybody a lot of prefill time.
+
+## Signed backends
+
+Every backend image merged by CI is now signed with [sigstore](https://www.sigstore.dev/)/cosign, keyless via Fulcio and Rekor, including each per-arch entry under the manifest list ([#9823](https://github.com/mudler/LocalAI/pull/9823)). It uses OCI 1.1 referrers rather than the legacy `:tag.sig` convention.
+
+On your side, verification runs against a policy that the gallery declares:
+
+```yaml
+verification:
+  issuer_regex: "^https://token\\.actions\\.githubusercontent\\.com$"
+  identity_regex: "^https://github\\.com/mudler/LocalAI/\\.github/workflows/backend_merge\\.yml@.*$"
+  not_before: "2026-05-22T00:00:00Z"
+```
+
+A few details that took some thinking.
+
+`not_before` is the revocation lever. Keyless Fulcio certificates are ephemeral, so there is nothing to revoke on the signing side. Revocation has to be policy side: move the date forward in the gallery YAML and every signature older than it stops validating.
+
+The TUF trusted root is cached process-wide, so installing ten backends from one gallery does one fetch instead of ten.
+
+Digest pinning closes the window between verifying and pulling, which is otherwise a TOCTOU you could drive a truck through.
+
+Strict mode is `--require-backend-integrity`, or `LOCALAI_REQUIRE_BACKEND_INTEGRITY=true`. It turns a missing policy or an empty SHA256 from a warning into a hard failure.
+
+Now the honest part: strict mode is opt-in and off by default, and until a gallery ships a `verification:` block, installs go through with a warning. The default `backend/index.yaml` does not have the blocks populated yet, that is the next step. So today this is machinery that works and is not yet enforcing much. Turn on strict mode in production once your gallery is populated, not before, or you will just break your own installs.
+
+## The prompt cache was off
+
+`llama-cpp` has a server-side prompt cache. LocalAI was not enabling it. So every agent turn, every coding-assistant call, every OpenAI-compatible CLI with a long system prompt, re-prefilled that whole prompt from scratch.
+
+On the reported workload, a repeated system prompt took 5 to 8 minutes per call before this change and seconds after it. Your numbers will depend on how long your prompt is and what hardware you are on.
+
+Two defaults flipped ([#9925](https://github.com/mudler/LocalAI/pull/9925), [#9951](https://github.com/mudler/LocalAI/pull/9951)):
+
+1. `kv_unified` is now `true` in `grpc-server.cpp`. The old `false` was silently force-disabling `cache_idle_slots` at server init, so the host prompt cache got allocated and then never written across requests. That is the one that actually explains the behaviour.
+2. `prompt_cache_all` defaults to `true` at the YAML layer, matching upstream llama.cpp's own default in `common.h`. The per-request `cache_prompt` knob is on out of the box.
+
+You can opt out with `options: ["kv_unified:false"]` or `prompt_cache_all: false`, and there are new keys (`cache_idle_slots`, `checkpoint_every_nt`) if you want to tune it. The model configuration docs got a worked example for the repeated-system-prompt case and an explanation of how `kv_unified`, `cache_ram` and `cache_idle_slots` interact, because they interact in ways that are not obvious.
+
+## Who is burning the GPU
+
+The usage page could tell you how many tokens were spent. It could not tell you who spent them ([#9920](https://github.com/mudler/LocalAI/pull/9920)).
+
+`usage_records` gained a `Source` column (`apikey`, `web`, `legacy`) plus the API key id and name, with an idempotent backfill of older rows on `InitDB`. The auth middleware passes the resolved key and the request source through, and usage middleware snapshots the key id and name at write time, so a key you revoke later still reads correctly in history (it renders as `(revoked)` rather than vanishing).
+
+Two new endpoints:
+
+```
+GET /api/auth/usage/sources        # your own
+GET /api/auth/admin/usage/sources  # everyone, with user_id / api_key_id filters
+```
+
+The admin view truncates at 200 keys. The React usage page gained a Sources tab with a source-mix ribbon, a top-7-plus-Other time chart, and a sortable table. Web interface session traffic is split per user instead of being lumped into one global row.
+
+## Distributed v3, and one good bug
+
+This one is worth writing down because the symptom and the cause were far apart.
+
+An operator reported this:
+
+```
+dgx-spark1     loaded   in_flight=6
+nvidia-thor1   loaded   in_flight=0
+```
+
+Two replicas of the same model, one taking everything, one idle forever. The round-robin was there and looked correct.
+
+The cause: `ModelLoader.Load` cached a `*Model` whose embedded `InFlightTrackingClient` was bound to a single `(nodeID, replicaIndex)`. The first request picked a node and got wrapped. Every request after that reused the wrapper, so it kept going to whichever node won the first pick, even after the reconciler scaled the model out. The routing code was fine. It just was not being consulted again!
+
+`SmartRouter.Route` now runs per request ([#9968](https://github.com/mudler/LocalAI/pull/9968)), the `in_flight ASC, last_used ASC, available_vram DESC` ordering actually fires, and replica selection lives in one place (`PickBestReplica`) with a spec asserting the SQL `ORDER BY` and the Go picker agree on a seeded dataset. `probeHealth` is memoized per `(nodeID, addr)` with a 30 second TTL and `singleflight` coalescing, because llama.cpp serializes `HealthCheck` against in-flight `Predict` and a burst of new requests would otherwise stall on it.
+
+Two other distributed changes.
+
+`POST /api/nodes/:id/backends/install` used to block for up to 3 minutes while the worker pulled the image, which froze the Backends picker in the interface. It returns HTTP 202 and a `jobID` immediately now ([#9928](https://github.com/mudler/LocalAI/pull/9928)). Install and upgrade timeouts are configurable via `LOCALAI_NATS_BACKEND_INSTALL_TIMEOUT` and `LOCALAI_NATS_BACKEND_UPGRADE_TIMEOUT`, defaulting to 15 minutes instead of the hardcoded 3. A NATS round-trip timeout while the worker is still pulling reports as `running_on_worker` rather than a hard failure.
+
+Workers also publish debounced install progress (~250ms) that the master forwards into the operations status ([#9958](https://github.com/mudler/LocalAI/pull/9958)), so distributed installs show per-byte progress the same way local ones do. Old workers stay silent and new masters tolerate the silence, so mixed-version clusters keep working.
+
+## Smaller things
+
+`LOCALAI_TRACING_MAX_BODY_BYTES` caps trace payload size, which stops the admin Traces page from trying to render a 40 MB embedding response.
+
+There is a `flake.nix` with a dev shell for NixOS users who do not want to go through Docker.
+
+The `vllm`, `sglang` and `vllm-omni` L4T13 backends are back for Jetson and DGX boxes, switched to PyPI aarch64+cu130 wheels to fix the torch 2.10 ABI mismatch.
+
+A distributed test harness landed in `tests/distributed/`, aimed at catching the class of regression the replica-pinning bug belonged to.
+
+## Thanks
+
+If you run LocalAI in production, the two things to look at here are strict mode (once your gallery has a `verification:` block) and whether the prompt cache change speeds up your workload. I would like to hear numbers from real setups, mine are one data point.
+
+[Full release notes](https://github.com/mudler/LocalAI/releases/tag/v4.3.0).

website/content/blog/why-we-write-our-own-engines.md

@@ -4,21 +4,21 @@ date: 2026-07-24
 author: "Ettore Di Giacinto"
 category: "Engineering"
 tags: ["engineering", "ggml", "vllm.cpp", "depth-anything.cpp", "parity"]
-summary: "A 66 MiB binary instead of a 9.1 GiB virtualenv, depth estimation that beats PyTorch on CPU in half the memory, and biometrics that match insightface bit for bit. The method, the measurements, and what it costs us."
+summary: "Eighteen of our backends are C or C++ ports we wrote from scratch instead of wrapping an upstream engine. Here is why we did it and what we measured."
 extracss: ["blog.css"]
 ---
 
-Most LocalAI backends wrap somebody else's engine, and that is the right default. llama.cpp, vLLM, whisper.cpp, stable-diffusion, MLX and the rest are maintained by people who are better at those models than we are, and wrapping them costs a Dockerfile and a gRPC shim.
+Most LocalAI backends wrap somebody else's engine. llama.cpp, vLLM, whisper.cpp, stable-diffusion and MLX are maintained by people who work on those models full time, and wrapping one of them costs us a Dockerfile and a gRPC shim. We do that wherever we can.
 
-Eighteen of our backends do not wrap anything. They are C or C++ ports we wrote from scratch, and each one exists because wrapping the upstream engine would have meant shipping something we could not ship: a multi-gigabyte Python install, a non-portable CUDA-only stack, or a model that had no C++ implementation at all. This post is about what those ports buy, measured, and what they cost.
+Eighteen of our backends do not wrap anything. They are C or C++ ports we wrote from scratch, and each one exists because wrapping the upstream engine would have meant shipping something we could not ship: a multi-gigabyte Python install, a CUDA-only stack that will not run on half the machines our users have, or, in a few cases, a model with no C++ implementation to wrap in the first place. Below are the numbers for four of them, and what keeping them alive takes.
 
-## What you get: one file, and memory you can predict
+## vllm.cpp: 66 MiB instead of 9.1 GiB
 
-Deploying a Python inference stack means resolving a dependency tree at install time, on the target machine, against whatever CUDA and glibc it has. Deploying a ggml port means copying a shared library and a GGUF file.
+Deploying a Python inference stack means resolving a dependency tree at install time, on the target machine, against whatever CUDA and glibc that machine has. Deploying a ggml port means copying a shared library and a GGUF file.
 
-The clearest measurement of that difference is [vllm.cpp](https://github.com/mudler/vllm.cpp), our C++20 port of vLLM's V1 serving architecture. Installing vLLM produces a 9.1 GiB virtualenv. Installing vllm.cpp produces a 66 MiB binary. The engine implements the same things the Python original does, including paged KV cache, continuous batching, prefix caching, the scheduler and the sampler, with no Python, no PyTorch and no ggml at inference.
+[vllm.cpp](https://github.com/mudler/vllm.cpp) is our C++20 port of vLLM's V1 serving architecture. Installing vLLM produces a 9.1 GiB virtualenv. Installing vllm.cpp produces a 66 MiB binary. It implements the same things the Python original does, including paged KV cache, continuous batching, prefix caching, the scheduler and the sampler, with no Python, no PyTorch and no ggml at inference.
 
-The obvious question is what that costs in throughput. On an NVIDIA GB10 running Qwen3.6-27B in NVFP4, greedy, closed loop, against vLLM in its production graphed configuration rather than `--enforce-eager`:
+The question is what that does to throughput. On an NVIDIA GB10 running Qwen3.6-27B in NVFP4, greedy, closed loop, against vLLM in its production graphed configuration rather than `--enforce-eager`:
 
 <div class="tw">
 <table>
@@ -31,13 +31,15 @@ The obvious question is what that costs in throughput. On an NVIDIA GB10 running
 </table>
 </div>
 
-We are ahead at all six points, and five of those six are ties. Our run-to-run noise band is 0.5%, and concurrency 2 through 32 land between 0.7% and 1.7%, so the honest reading is that only the single-stream case (4.5%) is clearly outside noise. Output is token-for-token identical to vLLM at every point on that curve. Peak host memory is 24.88 GiB against 28.18 GiB.
+Those are ties. Our run-to-run noise band is 0.5%, and concurrency 2 through 32 land between 0.7% and 1.7%, so those five points sit inside the noise or close enough to it not to matter. Only the single-stream case, at 4.5%, is clearly outside. Output is token-for-token identical to vLLM at every point on the curve, and peak host memory is 24.88 GiB against 28.18 GiB.
 
-A tie against a mature CUDA stack is a good result for a 66 MiB binary, and it means the footprint saving is not paid for in throughput. Against llama.cpp on CPU from the same GGUF file, prefill runs 1.18x faster (223.8 against 177.3 tok/s), decode is a tie inside llama.cpp's own spread, and the tokens are byte-identical to its greedy decode. Against MLX-LM on an Apple M4, prefill time to first token is 1.5% ahead and warm total throughput is 97.6% of MLX-LM, a real 2.4% gap that sits entirely in decode.
+The install drops from 9.1 GiB to 66 MiB and the throughput stays where it was, which is what we were after.
 
-## Sometimes the port is simply faster
+Against llama.cpp on CPU from the same GGUF file, prefill runs 1.18x faster (223.8 against 177.3 tok/s), decode is a tie inside llama.cpp's own spread, and the tokens match its greedy decode exactly. Against MLX-LM on an Apple M4, prefill time to first token is 1.5% ahead and warm total throughput is 97.6% of MLX-LM, a real 2.4% gap that sits entirely in decode.
 
-[depth-anything.cpp](https://github.com/mudler/depth-anything.cpp) is a port of ByteDance's Depth Anything 3, which gives you metric depth in metres from one ordinary photo, plus per-pixel confidence, camera intrinsics and extrinsics, and a back-projected point cloud. On CPU it is faster than PyTorch running the same model.
+## depth-anything.cpp is faster on CPU
+
+[depth-anything.cpp](https://github.com/mudler/depth-anything.cpp) is a port of ByteDance's Depth Anything 3, which gives you metric depth in metres from one ordinary photo, plus per-pixel confidence, camera intrinsics and extrinsics, and a back-projected point cloud. On CPU it runs faster than PyTorch on the same model.
 
 <div class="tw">
 <table>
@@ -49,40 +51,42 @@ A tie against a mature CUDA stack is a good result for a 66 MiB binary, and it m
 </table>
 </div>
 
-Same model, 1.31x the speed, 27% of the memory, and a load that finishes in 40 ms instead of 749 ms, on a Ryzen 9 9950X3D at 504x336 with 16 threads. The quantized q4_k build is a 99 MB file and stays near-lossless. Output correlates 1.0 with the reference forward pass, component by component, across 37 parity tests.
+That is on a Ryzen 9 9950X3D at 504x336 with 16 threads. The C++ build runs the same model 1.31x faster, uses 363 MB of RAM against 1328 MB, and loads in 40 ms instead of 749 ms. The quantized q4_k build is a 99 MB file and stays near-lossless. Output correlates 1.0 with the reference forward pass across 37 parity tests.
 
-The reason it is faster has nothing to do with writing better matmul kernels than PyTorch. Two positional embeddings, the DPT head's UV embedding and the backbone's bicubic position embedding, were being recomputed on every forward pass with single-threaded scalar sin, cos and bicubic loops, even though they depend only on the input geometry and are identical every call. Caching them removed about 95 ms of host-side overhead per forward, which is most of the gap. PyTorch builds the same embeddings with vectorized operations and never paid that cost.
+We did not write a better matmul kernel than PyTorch. Two positional embeddings, the DPT head's UV embedding and the backbone's bicubic position embedding, were being recomputed on every forward pass with single-threaded scalar sin, cos and bicubic loops, even though they depend only on the input geometry and are identical every call. Caching them removed about 95 ms of host-side overhead per forward, which is most of the gap. PyTorch builds the same embeddings with vectorized operations and never had that overhead to begin with.
 
-That is the general shape of these wins. The heavy GEMMs are close to a wash, because everyone is calling into the same class of BLAS kernel. The difference sits in host-side work that a Python reference implementation never bothered to optimize, and in not loading an interpreter and a framework to do inference. On GPU the picture flips back to parity: with the ggml CUDA backend and flash attention on a GB10, depth-anything.cpp ties PyTorch's tuned cuDNN at 47.3 ms per forward, and wins only the cold start, loading 1.75x to 2.9x faster.
+The heavy GEMMs are close to a wash, because everyone is calling into the same class of BLAS kernel. What is left is host-side work that a Python reference implementation never bothered to optimize, plus not loading an interpreter and a framework to do inference. On GPU it goes back to parity: with the ggml CUDA backend and flash attention on a GB10, depth-anything.cpp ties PyTorch's tuned cuDNN at 47.3 ms per forward, and wins only the cold start, loading 1.75x to 2.9x faster.
 
-## Parity is the gate, speed is the follow-up
+## The two where we are slower
 
-[face-detect.cpp](https://github.com/mudler/face-detect.cpp) and [voice-detect.cpp](https://github.com/localai-org/voice-detect.cpp) replaced LocalAI's Python `insightface` and `speaker-recognition` backends. Both are the case where we do not claim a CPU speed win, and both shipped anyway.
+[face-detect.cpp](https://github.com/mudler/face-detect.cpp) and [voice-detect.cpp](https://github.com/localai-org/voice-detect.cpp) replaced LocalAI's Python `insightface` and `speaker-recognition` backends. Neither of them is faster than what it replaced on CPU, and we shipped them anyway.
 
-face-detect.cpp runs the whole insightface buffalo chain, so SCRFD detection, five-landmark similarity-transform alignment to 112x112, and the ArcFace embedding, out of one self-contained GGUF with no Python and no onnxruntime. Detector boxes and landmarks match insightface to within 1 pixel, and the recognition embedding matches to cosine 1.000000, held at any thread count. On CPU it is slower than onnxruntime: SCRFD detect runs at about 0.83x at one thread and 0.69x at eight, ArcFace embed at about 0.61x and 0.84x. onnxruntime's MLAS convolution kernels sit at the FMA-port peak, and a custom AVX2 Winograd path narrowed the gap without closing it. On GPU, routing the same convolutions through cuDNN takes SCRFD from 14.8 ms to 6.4 ms and lands at torch-cuDNN parity.
+face-detect.cpp runs the whole insightface buffalo chain, so SCRFD detection, five-landmark similarity-transform alignment to 112x112, and the ArcFace embedding, out of one self-contained GGUF with no Python and no onnxruntime. Detector boxes and landmarks match insightface to within 1 pixel, and the recognition embedding matches to cosine 1.000000 at any thread count. On CPU it is slower than onnxruntime: SCRFD detect runs at about 0.83x at one thread and 0.69x at eight, ArcFace embed at about 0.61x and 0.84x. onnxruntime's MLAS convolution kernels sit at the FMA-port peak, and a custom AVX2 Winograd path narrowed the gap without closing it. On GPU, routing the same convolutions through cuDNN takes SCRFD from 14.8 ms to 6.4 ms, which lands at torch-cuDNN parity.
 
-voice-detect.cpp is the same story with a memory result attached. A WeSpeaker verification peaks at about 62 MB in our binary against about 334 MB for the CPU-only Python, torch and onnxruntime path, roughly 5.4x lower, with an identical verdict and embedding cosine 1.000000. End to end on CPU the two land within 10 to 15% of each other, trading the lead by model and thread count, and on GPU the conv encoders match the reference.
+voice-detect.cpp has a memory result instead. A WeSpeaker verification peaks at about 62 MB in our binary against about 334 MB for the CPU-only Python, torch and onnxruntime path, roughly 5.4x lower, with an identical verdict and embedding cosine 1.000000. End to end on CPU the two land within 10 to 15% of each other, trading the lead by model and thread count, and on GPU the conv encoders match the reference.
 
-For a biometric pipeline, matching the reference exactly matters more than being faster than it. An embedding that differs in the fourth decimal place changes verification decisions at a threshold, and every enrolled template in a deployment would have to be recomputed. Parity is what makes the replacement a drop-in rather than a migration.
+For a biometric pipeline we would rather have the exact match than the speed. An embedding that differs in the fourth decimal place changes verification decisions at a threshold, and every enrolled template in a deployment would have to be recomputed. Matching insightface exactly is what lets somebody swap the backend out without re-enrolling their users.
 
-## The method
+## How we do it
 
-Every port follows the same sequence, and the order is the important part.
+Every port follows the same four steps.
 
 Convert the weights first, into one GGUF with the tokenizer, the vocabulary and any auxiliary model embedded, so that deploying the model is copying a file.
 
-Port the graph second, and gate it component by component against reference tensors dumped from the original implementation. depth-anything.cpp has 37 ctest cases covering preprocessing, backbone, attention, the DPT head, depth, pose, the ray head, the ray to pose solver and the exporters. parakeet.cpp gates on transcript agreement with NeMo at WER 0. face-detect.cpp gates on box and landmark distance in pixels and embedding cosine. A port that is fast and slightly wrong is worthless, and without a per-component gate you find out it is wrong months later.
+Port the graph second, and check it component by component against reference tensors dumped from the original implementation. depth-anything.cpp has 37 ctest cases covering preprocessing, backbone, attention, the DPT head, depth, pose, the ray head, the ray to pose solver and the exporters. parakeet.cpp checks transcript agreement with NeMo at WER 0. face-detect.cpp checks box and landmark distance in pixels, and embedding cosine. Skip this step and you find out the port is wrong months later, from a user, on a model you had stopped thinking about.
 
-Optimize third, with a profiler, and only after parity holds. In parakeet.cpp the decisive win was caching a prediction-network LSTM forward pass that was 97% of transducer decode time and mostly redundant. In depth-anything.cpp it was two cached positional embeddings. Neither was a kernel rewrite, and neither would have been findable without a working baseline to profile.
+Optimize third, with a profiler, and only once the parity checks pass. In parakeet.cpp the win was caching a prediction-network LSTM forward pass that was 97% of transducer decode time and mostly redundant. In depth-anything.cpp it was the two positional embeddings above. Neither was a kernel rewrite, and neither would have turned up without a working baseline to profile.
 
 Expose a flat C ABI last. LocalAI dlopens the shared library through purego and calls that ABI directly, so there is no subprocess, no gRPC hop to a Python server, and no interpreter in the serving path.
 
-## What it costs
+## What it takes to maintain
 
-Maintenance, mostly. Each engine is a repository with its own CI, its own benchmark suite, its own GGUF conversion script and its own parity baselines, and upstream keeps releasing new checkpoints that need converter work.
+Each engine is its own repository with its own CI, benchmark suite, GGUF conversion script and parity baselines, and upstream keeps releasing checkpoints that need converter work.
 
 GPU kernels are the weak spot. ggml's generic CUDA convolution and attention kernels trail NVIDIA's tuned cuDNN on the conv-heavy models, which is why face-detect.cpp needs an explicit cuDNN path to reach parity, and why parakeet.cpp's GPU margin over NeMo is a median 1.25x while its CPU margin is wider.
 
-Porting also does not scale to everything. llama.cpp, vLLM, whisper.cpp, MLX and diffusers stay wrapped, because those projects are large, fast-moving and already excellent at what they do. We write an engine when a model has no C++ implementation, when the Python dependency is heavier than the model, or when the thing we need does not exist yet. Everything else we install from somebody else.
+It also does not scale to everything. llama.cpp, vLLM, whisper.cpp, MLX and diffusers stay wrapped, because those projects are large, fast-moving and already good at what they do. We write an engine when a model has no C++ implementation, when the Python dependency is heavier than the model itself, or when the thing we need does not exist yet. The rest we install like everybody else.
 
-Every engine listed above keeps its own benchmark suite, its parity gates and its methodology in its own repository, including the runs that did not work. The full list of them is the "Backends built by us" table in the [LocalAI README](https://github.com/mudler/LocalAI#backends-built-by-us).
+One thing that confuses people reading the tree for the first time: LocalAI's own core is Go, and each backend is written in whatever its model's ecosystem needs, which is why there is C++ sitting next to Python in the same repository.
+
+Every engine above keeps its benchmark suite, its parity checks and its methodology in its own repository, including the runs that did not work out. The full list is the "Backends built by us" table in the [LocalAI README](https://github.com/mudler/LocalAI#backends-built-by-us).

website/content/engines/_index.md

@@ -1,5 +1,5 @@
 ---
 title: "Engines"
-description: "Nineteen native C, C++ and Go engines written by the LocalAI team. No Python at inference, checked against the reference implementation in CI, and small enough to ship as one file."
+description: "Eighteen native C, C++ and Go engines written by the LocalAI team. No Python at inference, checked against the reference implementation in CI, and small enough to ship as one file."
 extracss: ["engines.css"]
 ---

website/data/engines.yaml

@@ -1,4 +1,4 @@
-# The nineteen native engines the LocalAI team wrote, and the one quantization
+# The eighteen native engines the LocalAI team wrote, and the one quantization
 # recipe that feeds them. This file is the single source of truth for the
 # /engines/ page: the layout renders whatever is here, in this order, and adds
 # nothing of its own. Numbers in `highlights` come from each engine's own

website/layouts/engines/list.html

@@ -10,7 +10,7 @@
     <p class="kicker fd" style="margin-top:0">Engines we build</p>
     <h1 class="eng-h1"><u><b>Eighteen engines,</b></u><u><b><s>written from scratch.</s></b></u></h1>
     <div class="bars" aria-hidden="true"><i></i><i></i><i></i><i></i></div>
-    <p class="lede fd mt2">Most backends wrap somebody else's engine. These do not. Each one exists because the thing we needed was a multi-gigabyte Python install, or closed, or nobody had built it yet. What you get instead is a binary and a GGUF file, checked against the reference implementation in CI, running on the machine you already own.</p>
+    <p class="lede fd mt2">Most LocalAI backends wrap somebody else's engine. These were written from scratch, each one because the thing we needed was a multi-gigabyte Python install, or closed, or nobody had built it yet. What ships instead is a binary and a GGUF file, checked against the reference implementation in CI, running on the machine you already own.</p>
     <div class="acts fd">
       <a class="btn" href="/#start">Install LocalAI <span>&#8594;</span></a>
       <a class="btn btn--o" href="/docs/features/backends/">How backends work</a>
@@ -88,8 +88,8 @@
   <div class="shell">
     <div class="bars rv" aria-hidden="true"><i></i><i></i><i></i><i></i></div>
     <p class="kicker rv">The rule we hold them to</p>
-    <h2 class="rv mt1" style="max-width:20ch">A port only ships once it matches the original.</h2>
-    <p class="lede rv mt2">Every engine here is gated against the framework it replaces, on the same input, on the same machine. That means a transcript that comes out word for word identical, boxes that land on the same pixels, or a waveform inside a stated tolerance. Speed is the part we then go and win, and the numbers on this page come out of each engine's own benchmark suite, not a marketing run.</p>
+    <h2 class="rv mt1" style="max-width:20ch">We do not ship a port until it matches the original.</h2>
+    <p class="lede rv mt2">Every engine here is gated against the framework it replaces, on the same input, on the same machine. That means a transcript identical to the reference, boxes that land on the same pixels, or a waveform inside a stated tolerance. Speed work comes after that, and the numbers on this page come out of each engine's own benchmark suite.</p>
     <div class="acts rv">
       <a class="btn" href="/#start">Install LocalAI &#8594;</a>
       <a class="btn btn--o" href="https://github.com/mudler/LocalAI">LocalAI on GitHub &#8599;</a>

website/layouts/index.html

@@ -37,7 +37,7 @@
   <div class="shell">
     <div class="bars rv" aria-hidden="true"><i></i><i></i><i></i><i></i></div>
     <p class="kicker rv">The runtime</p>
-    <h2 class="rv mt1" style="max-width:21ch">LocalAI is the engine everything else plugs into.</h2>
+    <h2 class="rv mt1" style="max-width:21ch">Everything else plugs into LocalAI.</h2>
     <p class="lede rv mt2">One binary with an OpenAI-compatible API in front of it. Point an existing client at it and the calls keep working, except now the model is on your machine. It also speaks the Anthropic, Ollama and ElevenLabs APIs, so most tools need a URL change and nothing else.</p>
     <p class="lede rv mt2">Underneath, a small core pulls each engine in as a separate backend, only when a model asks for it. That is why one install covers this much ground without becoming a 9 GB download.</p>
     <div class="apis rv">
@@ -75,7 +75,7 @@
       <div class="mi rv">
         <p class="mi__n">01 / HARDWARE</p>
         <h3>Every feature ships a CPU path first.</h3>
-        <p>Not a degraded mode that technically runs. The real one, tested in CI, on the hardware most people already have. GPUs make it faster, they are not the price of entry.</p>
+        <p>That path is tested in CI, on the hardware most people already have, and it is not a degraded fallback. A GPU makes it faster but is not required.</p>
         <p class="mi__meta">x86_64 · ARM64 · CUDA · ROCm · SYCL · Metal · Vulkan</p>
       </div>
       <div class="mi rv">
@@ -86,7 +86,7 @@
       </div>
       <div class="mi rv">
         <p class="mi__n">03 / DISTRIBUTED</p>
-        <h3>Plug in a second machine and stop there.</h3>
+        <h3>Add a second machine.</h3>
         <p>Routing, VRAM-aware placement, prefix-cache affinity and failover are the runtime's problem. You add hardware, the cluster works out what to do with it.</p>
         <p class="mi__meta">Smart routing · autoscaling · P2P · NATS · federation</p>
       </div>
@@ -174,7 +174,7 @@
       <div>
         <h3>parakeet.cpp</h3>
         <p class="spot__h">Twenty-seven times faster than whisper.cpp, on a CPU.</p>
-        <p>NVIDIA NeMo Parakeet, ported to C++ and ggml. Ten checkpoints, all of them verified at WER 0 against NeMo, which means the transcript comes out byte for byte identical while finishing first. Cache-aware streaming with end-of-utterance detection handles live audio, and the multilingual streaming model covers 40 or more locales.</p>
+        <p>NVIDIA NeMo Parakeet, ported to C++ and ggml. Ten checkpoints, all of them verified at WER 0 against NeMo, which means the transcript is identical to NeMo's while finishing first. Cache-aware streaming with end-of-utterance detection handles live audio, and the multilingual streaming model covers 40 or more locales.</p>
         <div class="facts">
           <div><b>27x</b><span>vs whisper.cpp, CPU</span></div>
           <div><b>1.40x</b><span>vs NeMo, CPU median</span></div>
@@ -395,7 +395,7 @@
           <p>Distributed mode with VRAM-aware routing, autoscaling, multi-user auth and per-user quotas.</p></div>
         <div class="tl__i"><p class="tl__d">MAY 2026</p><h4>It sees and hears</h4>
           <p>Voice recognition, face recognition with liveness, diarization, video generation, drop-in Ollama API.</p></div>
-        <div class="tl__i"><p class="tl__d">JUL 2026</p><h4>Nineteen engines of our own</h4>
+        <div class="tl__i"><p class="tl__d">JUL 2026</p><h4>Eighteen engines of our own</h4>
           <p>The native C and C++ ports take over the heavy Python backends, one modality at a time.</p></div>
       </div>
     </div>