@mlx-node/lm 0.0.12 → 0.0.15

package/dist/model-discovery.js

@@ -0,0 +1,274 @@
+/** Shared, native-free discovery of local chat checkpoints. */
+import { readdir, stat } from 'node:fs/promises';
+import { basename, join } from 'node:path';
+import { launchPresetFor, familyTraitsFor, NON_GENERATIVE_FAMILY_IDS, MODEL_FAMILY_DATA, } from './family-data.js';
+import { detectModelType, readGgufArchitecture, readModelConfig } from './model-detection.js';
+/**
+ * The Qwen3.5/Qwen3.8 discovery filter retains its XL policy. Gemma4 and Muse
+ * accept all supported tensor formats, including Q4_0 QAT checkpoints.
+ * Match the Unsloth Dynamic XL target names users download, while excluding
+ * ordinary Q4_K_M files and companion artifacts such as imatrix/mmproj/draft.
+ */
+const QWEN35_XL_GGUF = /(?:^|[-_.])Q\d+_K_XL\.gguf$/i;
+const GGUF_COMPANION_NAME = /(?:^|[-_.])(?:imatrix|mmproj|dflash|draft)(?:[-_.]|$)/i;
+// Match the native loaders' primary files/shards. A draft or projector
+// SafeTensors file beside a GGUF is not a converted target checkpoint.
+const PRIMARY_SAFETENSORS = /^(?:model|weights)\.safetensors$|^model(?:-|\.safetensors-).+-of-.+\.safetensors$/;
+function isQwen35XlGguf(name) {
+    return QWEN35_XL_GGUF.test(name) && !GGUF_COMPANION_NAME.test(name);
+}
+function ggufModelName(name) {
+    return name.slice(0, -'.gguf'.length);
+}
+function requiresGgufAssets(modelType) {
+    return modelType === 'gemma4' || modelType === 'muse_glimmer';
+}
+function matchesGgufFamily(path, modelType) {
+    const architecture = readGgufArchitecture(path);
+    return MODEL_FAMILY_DATA.some((family) => family.id === modelType &&
+        'ggufArchitectures' in family &&
+        family.ggufArchitectures.some((supported) => supported === architecture));
+}
+async function hasGgufAssets(modelDir) {
+    try {
+        const assets = await Promise.all(['config.json', 'tokenizer.json'].map((name) => stat(join(modelDir, name))));
+        return assets.every((asset) => asset.isFile());
+    }
+    catch {
+        return false;
+    }
+}
+async function modelFileInventory(modelDir) {
+    try {
+        const files = (await readdir(modelDir, { withFileTypes: true }))
+            .filter((entry) => entry.isFile())
+            .map((entry) => entry.name);
+        return {
+            xlGgufs: files.filter(isQwen35XlGguf).sort(),
+            targetGgufs: files
+                .filter((name) => name.toLowerCase().endsWith('.gguf') && !GGUF_COMPANION_NAME.test(name))
+                .sort(),
+            hasGguf: files.some((name) => name.toLowerCase().endsWith('.gguf')),
+            hasSafetensors: files.some((name) => name.toLowerCase().endsWith('.safetensors')),
+            hasPrimarySafetensors: files.some((name) => PRIMARY_SAFETENSORS.test(name)),
+        };
+    }
+    catch {
+        return { xlGgufs: [], targetGgufs: [], hasGguf: false, hasSafetensors: false, hasPrimarySafetensors: false };
+    }
+}
+function positiveInteger(value) {
+    return typeof value === 'number' && Number.isFinite(value) && value > 0 ? Math.floor(value) : undefined;
+}
+function nonEmptyRecord(value) {
+    return typeof value === 'object' && value !== null && !Array.isArray(value) && Object.keys(value).length > 0;
+}
+/**
+ * Read cheap discovery metadata from `<modelPath>/config.json`.
+ *
+ * The trained context window comes from:
+ * root `max_position_embeddings` first (qwen3, lfm2), then
+ * `text_config.max_position_embeddings` (qwen3_5, qwen3_5_moe, gemma4
+ * unified), else the family fallback.
+ *
+ * Image support is advertised only when a family with a native multimodal
+ * implementation carries its valid, non-empty vision marker: `vision_config`
+ * for Qwen, and either `vision_config` or `unified_vision_config` for Gemma.
+ * This lets Pi's model picker and `--list-models` expose checkpoint capability
+ * without loading weights. The first resident load remains authoritative and
+ * reconciles this optimistic config-level advertisement via
+ * `session.supportsImages()` (for example, when conversion stripped an
+ * incompatible vision tower).
+ *
+ * `detectModelType` already parsed this file, so a read/parse failure here
+ * (e.g. a racing rewrite) lands on the context fallback and text-only input
+ * instead of dropping the model or guessing a positive capability.
+ */
+async function readDiscoveryMetadata(modelPath, modelType, fallbackContextWindow) {
+    try {
+        const config = (await readModelConfig(modelPath));
+        const root = positiveInteger(config.max_position_embeddings);
+        const textConfig = config.text_config;
+        const nested = nonEmptyRecord(textConfig) ? positiveInteger(textConfig.max_position_embeddings) : undefined;
+        const hasVisionConfig = nonEmptyRecord(config.vision_config);
+        const supportsImages = modelType === 'gemma4'
+            ? hasVisionConfig || nonEmptyRecord(config.unified_vision_config)
+            : (modelType === 'qwen3_5' || modelType === 'qwen3_5_moe') && hasVisionConfig;
+        const draftOnly = Array.isArray(config.architectures) && config.architectures.includes('DFlash2DraftModel');
+        return {
+            contextWindow: root ?? nested ?? fallbackContextWindow,
+            supportsImages,
+            draftOnly,
+        };
+    }
+    catch {
+        return { contextWindow: fallbackContextWindow, supportsImages: false, draftOnly: false };
+    }
+}
+/**
+ * Scan `modelsDir` for chat-capable model subdirectories, Gemma4/Muse GGUFs, and
+ * dense Qwen3.5/Qwen3.8 `Q<number>_K_XL.gguf` files. GGUF files may live directly
+ * under `modelsDir` or one level inside a downloaded GGUF repository. Each is
+ * registered by filename stem so quant variants remain independently selectable.
+ *
+ * An unreadable dir yields `[]`. Entries with an undetectable config, a
+ * non-generative type, or no launch preset are skipped silently (warnings only
+ * when `MLX_DEBUG` is set). No weights are loaded. Results are sorted by name.
+ */
+export async function discoverLocalChatModels(modelsDir) {
+    const debug = Boolean(process.env.MLX_DEBUG);
+    let entries;
+    try {
+        entries = await readdir(modelsDir, { withFileTypes: true });
+    }
+    catch {
+        return [];
+    }
+    // Collision resolution below gives the first occurrence the bare filename
+    // stem. Directory enumeration order is unspecified, so sort before assigning
+    // IDs to keep persisted `mlx/<id>` selections stable across filesystems.
+    entries.sort((a, b) => (a.name < b.name ? -1 : a.name > b.name ? 1 : 0));
+    const out = [];
+    const usedNames = new Set();
+    const append = async (preferredName, path, metadataRoot, modelType, scopeName) => {
+        if (NON_GENERATIVE_FAMILY_IDS.has(modelType))
+            return;
+        // Fail-closed guards: dead-by-construction for chat families (the
+        // family-data row type requires traits + a preset), live for any foreign
+        // string that slips through detection.
+        const preset = launchPresetFor(modelType);
+        if (!preset) {
+            if (debug)
+                console.warn(`[mlx] skip ${path}: no launch preset for ${modelType}`);
+            return;
+        }
+        const traits = familyTraitsFor(modelType);
+        if (!traits) {
+            if (debug)
+                console.warn(`[mlx] skip ${path}: no FAMILY_TRAITS entry for ${modelType}`);
+            return;
+        }
+        const metadata = await readDiscoveryMetadata(metadataRoot, modelType, traits.fallbackContextWindow);
+        if (metadata.draftOnly) {
+            if (debug)
+                console.warn(`[mlx] skip ${path}: companion draft checkpoint is not a chat model`);
+            return;
+        }
+        let name = preferredName;
+        if (usedNames.has(name)) {
+            name = `${scopeName}-${preferredName}`;
+            let suffix = 2;
+            while (usedNames.has(name))
+                name = `${scopeName}-${preferredName}-${suffix++}`;
+        }
+        usedNames.add(name);
+        // Automatic companions can be installed or removed after this startup
+        // scan. The host resolves them on each load; draftModelPath is reserved
+        // for caller-supplied paths that the loader must treat as authoritative.
+        out.push({
+            name,
+            path,
+            modelType,
+            preset,
+            traits,
+            contextWindow: metadata.contextWindow,
+            supportsImages: metadata.supportsImages,
+        });
+    };
+    for (const entry of entries) {
+        if (entry.isFile() && entry.name.toLowerCase().endsWith('.gguf') && !GGUF_COMPANION_NAME.test(entry.name)) {
+            const full = join(modelsDir, entry.name);
+            try {
+                const modelType = await detectModelType(full);
+                // A shared sibling config can describe another target or a projector.
+                // Never advertise a file under a loader that disagrees with its header.
+                if (!matchesGgufFamily(full, modelType))
+                    continue;
+                if (requiresGgufAssets(modelType) && !(await hasGgufAssets(modelsDir))) {
+                    if (debug)
+                        console.warn(`[mlx] skip ${full}: native ${modelType} GGUF requires sibling config.json and tokenizer.json`);
+                    continue;
+                }
+                if (modelType === 'gemma4' ||
+                    modelType === 'muse_glimmer' ||
+                    (modelType === 'qwen3_5' && isQwen35XlGguf(entry.name))) {
+                    await append(ggufModelName(entry.name), full, modelsDir, modelType, basename(modelsDir));
+                }
+                else if (debug) {
+                    console.warn(`[mlx] skip ${full}: no supported direct GGUF target for ${modelType}`);
+                }
+            }
+            catch (err) {
+                if (debug)
+                    console.warn(`[mlx] skip ${full}: ${err.message}`);
+            }
+            continue;
+        }
+        if (!entry.isDirectory())
+            continue;
+        const full = join(modelsDir, entry.name);
+        let modelType;
+        try {
+            modelType = await detectModelType(full);
+        }
+        catch (err) {
+            if (debug)
+                console.warn(`[mlx] skip ${full}: ${err.message}`);
+            continue;
+        }
+        const inventory = await modelFileInventory(full);
+        const hasModelWeights = requiresGgufAssets(modelType) ? inventory.hasPrimarySafetensors : inventory.hasSafetensors;
+        if (requiresGgufAssets(modelType) && !hasModelWeights && inventory.targetGgufs.length > 0) {
+            if (!(await hasGgufAssets(full))) {
+                if (debug)
+                    console.warn(`[mlx] skip ${full}: native ${modelType} GGUF requires sibling config.json and tokenizer.json`);
+                continue;
+            }
+            for (const gguf of inventory.targetGgufs) {
+                const path = join(full, gguf);
+                try {
+                    if (!matchesGgufFamily(path, modelType))
+                        continue;
+                    await append(ggufModelName(gguf), path, full, modelType, entry.name);
+                }
+                catch (err) {
+                    if (debug)
+                        console.warn(`[mlx] skip ${path}: ${err.message}`);
+                }
+            }
+            continue;
+        }
+        const { xlGgufs } = inventory;
+        if (xlGgufs.length > 0 && !inventory.hasPrimarySafetensors) {
+            if (modelType !== 'qwen3_5') {
+                if (debug) {
+                    console.warn(`[mlx] skip ${full}: direct XL GGUF loading is not supported for ${modelType}`);
+                }
+                continue;
+            }
+            for (const gguf of xlGgufs) {
+                const path = join(full, gguf);
+                try {
+                    if (matchesGgufFamily(path, modelType))
+                        await append(ggufModelName(gguf), path, full, modelType, entry.name);
+                }
+                catch (err) {
+                    if (debug)
+                        console.warn(`[mlx] skip ${path}: ${err.message}`);
+                }
+            }
+            continue;
+        }
+        // Present each supported GGUF variant separately in the picker. Keep
+        // converted model directories discoverable when they retain
+        // an imatrix/source GGUF beside their actual SafeTensors weights.
+        if (inventory.hasGguf && !hasModelWeights) {
+            if (debug)
+                console.warn(`[mlx] skip ${full}: no supported direct GGUF target`);
+            continue;
+        }
+        await append(basename(full), full, full, modelType, entry.name);
+    }
+    out.sort((a, b) => (a.name < b.name ? -1 : a.name > b.name ? 1 : 0));
+    return out;
+}

package/dist/models/model-loader.d.ts

@@ -7,6 +7,8 @@ import { ChatSession, type SessionCapableModel } from '../chat-session.js';
 import { type ModelType, type TrainableFamilyId } from '../family-data.js';
 /** Optional settings for {@link loadModel} / {@link loadSession}. */
 export interface LoadModelOptions {
+    /** Discover a Qwen DFlash2 companion on disk (default true). Explicit draftModelPath wins. */
+    autoLoadDraft?: boolean;
     /**
      * Directory of an external draft checkpoint (config.json +
      * model.safetensors) loaded alongside the target for speculative decoding.
@@ -125,6 +127,8 @@ export type TrainableModel = Awaited<ReturnType<LoaderBindings[TrainableFamilyId
  * rejects it.
  * Without the option, Gemma4 loads `<modelPath>/draft/` automatically when
  * that embedded checkpoint is present.
+ * Qwen also discovers `draft/` or a shared `qwen3.8-27b-dflash2` directory
+ * beside a Qwen3.8-27B target. Set `autoLoadDraft: false` to disable discovery.
  */
 export declare function loadModel(modelPath: string, options?: LoadModelOptions): Promise<LoadableModel>;
 /**
@@ -148,6 +152,8 @@ export declare function loadModel(modelPath: string, options?: LoadModelOptions)
  * rejects it.
  * Without the option, Gemma4 loads `<modelPath>/draft/` automatically when
  * that embedded checkpoint is present.
+ * Qwen uses the same companion discovery as `loadModel()` unless
+ * `autoLoadDraft: false` is passed.
  * The resulting session auto-enables the speculative path when the model
  * reports `hasMtpWeights()` AND does not opt out of the auto-default; pass
  * `enableMtp: false` per call to suppress it, or `enableMtp: true` to force

package/dist/models/model-loader.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"model-loader.d.ts","sourceRoot":"","sources":["../../src/models/model-loader.ts"],"names":[],"mappings":"AAAA;;;GAGG;AAKH,OAAO,EACL,WAAW,IAAI,iBAAiB,EAEhC,YAAY,EACZ,SAAS,IAAI,eAAe,EAC5B,gBAAgB,IAAI,sBAAsB,EAC1C,cAAc,IAAI,oBAAoB,EACtC,eAAe,EACf,UAAU,IAAI,gBAAgB,EAC9B,WAAW,IAAI,iBAAiB,EAChC,cAAc,IAAI,oBAAoB,EACvC,MAAM,gBAAgB,CAAC;AAExB,OAAO,EAAE,WAAW,EAAE,KAAK,mBAAmB,EAAE,MAAM,oBAAoB,CAAC;AAC3E,OAAO,EAML,KAAK,SAAS,EACd,KAAK,iBAAiB,EACvB,MAAM,mBAAmB,CAAC;AAW3B,qEAAqE;AACrE,MAAM,WAAW,gBAAgB;IAC/B;;;;;;;;;;;;;;;;;OAiBG;IACH,cAAc,CAAC,EAAE,MAAM,CAAC;CACzB;AAmBD;;;;;;GAMG;AACH,QAAA,MAAM,eAAe;;iBAEjB,IAAI,cAAc,MAAM,YAAY,gBAAgB;;;iBAKpD,gBAAgB;;;iBAGhB,IAAI,cAAc,MAAM;;;iBACxB,gBAAgB;;;iBAGhB,IAAI,cAAc,MAAM;iBACxB,gBAAgB;;;iBAGhB,IAAI,cAAc,MAAM;;;;iBACxB,gBAAgB;;;iBAGhB,IAAI,cAAc,MAAM,YAAY,gBAAgB;;;iBAKpD,gBAAgB;;;iBAGhB,IAAI,cAAc,MAAM;;;iBACxB,gBAAgB;;;iBAGhB,IAAI,cAAc,MAAM;;;;iBACxB,gBAAgB;;;iBAGhB,IAAI,cAAc,MAAM;;;;iBACxB,gBAAgB;;;iBAGhB,IAAI,cAAc,MAAM;;;iBACxB,gBAAgB;;;iBAGhB,IAAI,cAAc,MAAM;iBACxB,gBAAgB;;;iBAGhB,IAAI,cAAc,MAAM;iBACxB,gBAAgB;;CAEiC,CAAC;AAEtD,KAAK,cAAc,GAAG,OAAO,eAAe,CAAC;AAE7C;;;;;;;GAOG;AACH,MAAM,MAAM,aAAa,GAAG,YAAY,CAAC,cAAc,CAAC,SAAS,CAAC,CAAC,kBAAkB,CAAC,CAAC,CAAC;AAExF;;;GAGG;AACH,MAAM,MAAM,cAAc,GACtB,OAAO,CAAC,UAAU,CAAC,cAAc,CAAC,iBAAiB,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,GAC9D,YAAY,CAAC,cAAc,CAAC,iBAAiB,CAAC,CAAC,kBAAkB,CAAC,CAAC,CAAC;AAsCxE;;;;;;;;;;;GAWG;AACH,wBAAsB,SAAS,CAAC,SAAS,EAAE,MAAM,EAAE,OAAO,CAAC,EAAE,gBAAgB,GAAG,OAAO,CAAC,aAAa,CAAC,CAGrG;AAED;;;;;;;;;;;;;;;;;;;;;;;;;;;GA2BG;AACH,wBAAsB,WAAW,CAC/B,SAAS,EAAE,MAAM,EACjB,OAAO,CAAC,EAAE,gBAAgB,GACzB,OAAO,CAAC,WAAW,CAAC,mBAAmB,CAAC,CAAC,CAa3C;AAED,wBAAsB,eAAe,CAAC,SAAS,EAAE,MAAM,GAAG,OAAO,CAAC,SAAS,CAAC,CAwB3E"}
+{"version":3,"file":"model-loader.d.ts","sourceRoot":"","sources":["../../src/models/model-loader.ts"],"names":[],"mappings":"AAAA;;;GAGG;AAEH,OAAO,EACL,WAAW,IAAI,iBAAiB,EAEhC,YAAY,EACZ,SAAS,IAAI,eAAe,EAC5B,gBAAgB,IAAI,sBAAsB,EAC1C,cAAc,IAAI,oBAAoB,EACtC,eAAe,EACf,UAAU,IAAI,gBAAgB,EAC9B,WAAW,IAAI,iBAAiB,EAChC,cAAc,IAAI,oBAAoB,EACvC,MAAM,gBAAgB,CAAC;AAExB,OAAO,EAAE,WAAW,EAAE,KAAK,mBAAmB,EAAE,MAAM,oBAAoB,CAAC;AAE3E,OAAO,EAAiB,KAAK,SAAS,EAAE,KAAK,iBAAiB,EAAE,MAAM,mBAAmB,CAAC;AAY1F,qEAAqE;AACrE,MAAM,WAAW,gBAAgB;IAC/B,8FAA8F;IAC9F,aAAa,CAAC,EAAE,OAAO,CAAC;IACxB;;;;;;;;;;;;;;;;;OAiBG;IACH,cAAc,CAAC,EAAE,MAAM,CAAC;CACzB;AAmBD;;;;;;GAMG;AACH,QAAA,MAAM,eAAe;;iBAEjB,IAAI,cAAc,MAAM,YAAY,gBAAgB;;;iBAKpD,gBAAgB;;;iBAGhB,IAAI,cAAc,MAAM;;;iBACxB,gBAAgB;;;iBAGhB,IAAI,cAAc,MAAM;iBACxB,gBAAgB;;;iBAGhB,IAAI,cAAc,MAAM;;;;iBACxB,gBAAgB;;;iBAGhB,IAAI,cAAc,MAAM,YAAY,gBAAgB;;;iBAKpD,gBAAgB;;;iBAGhB,IAAI,cAAc,MAAM;;;iBACxB,gBAAgB;;;iBAGhB,IAAI,cAAc,MAAM;;;;iBACxB,gBAAgB;;;iBAGhB,IAAI,cAAc,MAAM;;;;iBACxB,gBAAgB;;;iBAGhB,IAAI,cAAc,MAAM;;;iBACxB,gBAAgB;;;iBAGhB,IAAI,cAAc,MAAM;iBACxB,gBAAgB;;;iBAGhB,IAAI,cAAc,MAAM;iBACxB,gBAAgB;;CAEiC,CAAC;AAEtD,KAAK,cAAc,GAAG,OAAO,eAAe,CAAC;AAE7C;;;;;;;GAOG;AACH,MAAM,MAAM,aAAa,GAAG,YAAY,CAAC,cAAc,CAAC,SAAS,CAAC,CAAC,kBAAkB,CAAC,CAAC,CAAC;AAExF;;;GAGG;AACH,MAAM,MAAM,cAAc,GACtB,OAAO,CAAC,UAAU,CAAC,cAAc,CAAC,iBAAiB,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,GAC9D,YAAY,CAAC,cAAc,CAAC,iBAAiB,CAAC,CAAC,kBAAkB,CAAC,CAAC,CAAC;AAgCxE;;;;;;;;;;;;;GAaG;AACH,wBAAsB,SAAS,CAAC,SAAS,EAAE,MAAM,EAAE,OAAO,CAAC,EAAE,gBAAgB,GAAG,OAAO,CAAC,aAAa,CAAC,CAGrG;AAED;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA6BG;AACH,wBAAsB,WAAW,CAC/B,SAAS,EAAE,MAAM,EACjB,OAAO,CAAC,EAAE,gBAAgB,GACzB,OAAO,CAAC,WAAW,CAAC,mBAAmB,CAAC,CAAC,CAa3C;AAED,wBAAsB,eAAe,CAAC,SAAS,EAAE,MAAM,GAAG,OAAO,CAAC,SAAS,CAAC,CAE3E"}

package/dist/models/model-loader.js

@@ -2,11 +2,11 @@
  * Native half of the family registry: one loader binding per
  * `MODEL_FAMILY_DATA` row, plus `detectModelType` (filesystem + GGUF).
  */
-import { readFile } from 'node:fs/promises';
-import { dirname, extname, join } from 'node:path';
 import { Gemma4Model as NativeGemma4Model, ggufArchitecture, HarrierModel, Lfm2Model as NativeLfm2Model, MuseGlimmerModel as NativeMuseGlimmerModel, NemotronHModel as NativeNemotronHModel, QianfanOCRModel, Qwen3Model as NativeQwen3Model, Qwen35Model as NativeQwen35Model, Qwen35MoeModel as NativeQwen35MoeModel, } from '@mlx-node/core';
 import { ChatSession } from '../chat-session.js';
-import { familyDataFor, MalformedModelConfigError, matchFamily, MODEL_FAMILY_DATA, UnsupportedModelTypeError, } from '../family-data.js';
+import { findDFlash2Draft } from '../draft-companion.js';
+import { familyDataFor } from '../family-data.js';
+import { detectModelType as detectLocalModelType } from '../model-detection.js';
 import { Gemma4Model, Lfm2Model, MuseGlimmerModel, NemotronHModel, Qwen3Model, Qwen35Model, Qwen35MoeModel, } from '../stream.js';
 /**
  * Native half of the family registry: one loader + native class per
@@ -61,9 +61,6 @@ const LOADER_BINDINGS = {
         nativeModelClass: QianfanOCRModel,
     },
 };
-// Only families whose native `load(path)` accepts a GGUF file carry a
-// `ggufArchitectures` row entry (see family-data.ts).
-const GGUF_ARCHITECTURE_MODEL_TYPES = new Map(MODEL_FAMILY_DATA.flatMap((row) => 'ggufArchitectures' in row ? row.ggufArchitectures.map((architecture) => [architecture, row.id]) : []));
 function requireFamilyData(modelType) {
     const family = familyDataFor(modelType);
     if (family === undefined) {
@@ -82,7 +79,8 @@ function dispatchLoad(modelType, modelPath, options) {
             `${modelPath} has model_type "${modelType}"`);
     }
     const binding = LOADER_BINDINGS[modelType];
-    return binding.load(modelPath, options);
+    const draftModelPath = options?.draftModelPath ?? (options?.autoLoadDraft === false ? undefined : findDFlash2Draft(modelPath, modelType));
+    return binding.load(modelPath, draftModelPath === undefined ? options : { ...options, draftModelPath });
 }
 /**
  * Load a model from disk, auto-detecting architecture from config.json.
@@ -95,6 +93,8 @@ function dispatchLoad(modelType, modelPath, options) {
  * rejects it.
  * Without the option, Gemma4 loads `<modelPath>/draft/` automatically when
  * that embedded checkpoint is present.
+ * Qwen also discovers `draft/` or a shared `qwen3.8-27b-dflash2` directory
+ * beside a Qwen3.8-27B target. Set `autoLoadDraft: false` to disable discovery.
  */
 export async function loadModel(modelPath, options) {
     const modelType = await detectModelType(modelPath);
@@ -121,6 +121,8 @@ export async function loadModel(modelPath, options) {
  * rejects it.
  * Without the option, Gemma4 loads `<modelPath>/draft/` automatically when
  * that embedded checkpoint is present.
+ * Qwen uses the same companion discovery as `loadModel()` unless
+ * `autoLoadDraft: false` is passed.
  * The resulting session auto-enables the speculative path when the model
  * reports `hasMtpWeights()` AND does not opt out of the auto-default; pass
  * `enableMtp: false` per call to suppress it, or `enableMtp: true` to force
@@ -141,29 +143,5 @@ export async function loadSession(modelPath, options) {
     return new ChatSession(m);
 }
 export async function detectModelType(modelPath) {
-    const isGguf = extname(modelPath).toLowerCase() === '.gguf';
-    const configPath = isGguf ? join(dirname(modelPath), 'config.json') : join(modelPath, 'config.json');
-    let raw;
-    try {
-        raw = await readFile(configPath, 'utf-8');
-    }
-    catch (e) {
-        if (isGguf && typeof e === 'object' && e !== null && 'code' in e && e.code === 'ENOENT') {
-            const architecture = ggufArchitecture(modelPath);
-            const modelType = GGUF_ARCHITECTURE_MODEL_TYPES.get(architecture);
-            if (modelType === undefined) {
-                throw new Error(`Unsupported GGUF architecture "${architecture}" in ${modelPath}`);
-            }
-            return modelType;
-        }
-        throw new Error(`Cannot detect model type: config.json not found in ${modelPath}`);
-    }
-    try {
-        return matchFamily(modelPath, JSON.parse(raw));
-    }
-    catch (e) {
-        if (e instanceof UnsupportedModelTypeError || e instanceof MalformedModelConfigError)
-            throw e;
-        throw new Error(`Cannot detect model type: config.json not found in ${modelPath}`);
-    }
+    return detectLocalModelType(modelPath, ggufArchitecture);
 }

package/dist/models/paged-config-override.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"paged-config-override.d.ts","sourceRoot":"","sources":["../../src/models/paged-config-override.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAQH,iEAAiE;AACjE,eAAO,MAAM,wBAAwB,YAAI,SAAS,EAAE,aAAa,CAAU,CAAC;AAQ5E,MAAM,WAAW,iCAAiC;IAChD,qFAAqF;IACrF,UAAU,CAAC,EAAE,SAAS,MAAM,EAAE,CAAC;IAC/B,0EAA0E;IAC1E,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB;;;;OAIG;IACH,0BAA0B,CAAC,EAAE,OAAO,CAAC;CACtC;AAED;;;;;;;GAOG;AACH,qBAAa,0BAA0B;IACrC,OAAO,CAAC,QAAQ,CAAC,UAAU,CAAsB;IACjD,OAAO,CAAC,QAAQ,CAAC,aAAa,CAAS;IACvC,OAAO,CAAC,QAAQ,CAAC,0BAA0B,CAAU;IACrD,OAAO,CAAC,QAAQ,CAAC,SAAS,CAAsC;IAChE,OAAO,CAAC,QAAQ,CAAC,cAAc,CAA8B;IAC7D,OAAO,CAAC,WAAW,CAA8B;IACjD,OAAO,CAAC,cAAc,CAA4B;IAClD,OAAO,CAAC,QAAQ,CAAS;IAEzB,YAAY,OAAO,GAAE,iCAAsC,EAI1D;IAED;;;;;;;;;;;;;;;;;OAiBG;IACG,OAAO,CAAC,SAAS,EAAE,MAAM,EAAE,kBAAkB,CAAC,EAAE,MAAM,EAAE,iBAAiB,CAAC,EAAE,OAAO,GAAG,OAAO,CAAC,MAAM,CAAC,CAY1G;YAEa,eAAe;IAgF7B,6EAA6E;IAC7E,OAAO,IAAI,OAAO,CAAC,IAAI,CAAC,CAKvB;YAEa,cAAc;YAad,cAAc;IA+D5B,OAAO,CAAC,OAAO;CAIhB"}
+{"version":3,"file":"paged-config-override.d.ts","sourceRoot":"","sources":["../../src/models/paged-config-override.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAUH,iEAAiE;AACjE,eAAO,MAAM,wBAAwB,YAAI,SAAS,EAAE,aAAa,CAAU,CAAC;AAQ5E,MAAM,WAAW,iCAAiC;IAChD,qFAAqF;IACrF,UAAU,CAAC,EAAE,SAAS,MAAM,EAAE,CAAC;IAC/B,0EAA0E;IAC1E,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB;;;;OAIG;IACH,0BAA0B,CAAC,EAAE,OAAO,CAAC;CACtC;AAED;;;;;;;GAOG;AACH,qBAAa,0BAA0B;IACrC,OAAO,CAAC,QAAQ,CAAC,UAAU,CAAsB;IACjD,OAAO,CAAC,QAAQ,CAAC,aAAa,CAAS;IACvC,OAAO,CAAC,QAAQ,CAAC,0BAA0B,CAAU;IACrD,OAAO,CAAC,QAAQ,CAAC,SAAS,CAAsC;IAChE,OAAO,CAAC,QAAQ,CAAC,cAAc,CAA8B;IAC7D,OAAO,CAAC,WAAW,CAA8B;IACjD,OAAO,CAAC,cAAc,CAA4B;IAClD,OAAO,CAAC,QAAQ,CAAS;IAEzB,YAAY,OAAO,GAAE,iCAAsC,EAI1D;IAED;;;;;;;;;;;;;;;;;OAiBG;IACG,OAAO,CAAC,SAAS,EAAE,MAAM,EAAE,kBAAkB,CAAC,EAAE,MAAM,EAAE,iBAAiB,CAAC,EAAE,OAAO,GAAG,OAAO,CAAC,MAAM,CAAC,CAY1G;YAEa,eAAe;IAgG7B,6EAA6E;IAC7E,OAAO,IAAI,OAAO,CAAC,IAAI,CAAC,CAKvB;YAEa,cAAc;YAad,cAAc;IA+D5B,OAAO,CAAC,OAAO;CAIhB"}

package/dist/models/paged-config-override.js

@@ -11,7 +11,8 @@
  */
 import { mkdir, mkdtemp, readFile, readdir, rm, stat, symlink, writeFile } from 'node:fs/promises';
 import { tmpdir } from 'node:os';
-import { dirname, isAbsolute, join, relative, resolve, sep } from 'node:path';
+import { dirname, extname, isAbsolute, join, relative, resolve, sep } from 'node:path';
+import { ggufArchitecture, prepareMuseGlimmerGguf } from '@mlx-node/core';
 import { CHAT_FAMILY_IDS } from '../family-data.js';
 /** Families historically forced paged by `mlx launch claude`. */
 export const QWEN35_PAGED_MODEL_TYPES = ['qwen3_5', 'qwen3_5_moe'];
@@ -75,6 +76,25 @@ export class PagedConfigOverrideManager {
     }
     async resolveInternal(modelPath, canonicalModelType, persistPagedCache) {
         const sourcePath = isAbsolute(modelPath) ? modelPath : resolve(modelPath);
+        if (extname(sourcePath).toLowerCase() === '.gguf') {
+            let modelType = canonicalModelType;
+            if (modelType === undefined) {
+                try {
+                    if (ggufArchitecture(sourcePath) === 'muse-glimmer')
+                        modelType = 'muse_glimmer';
+                }
+                catch {
+                    return modelPath;
+                }
+            }
+            if (modelType !== 'muse_glimmer' || !this.modelTypes.has(modelType))
+                return modelPath;
+            // A symlink to the GGUF would canonicalize back to the original config.
+            // Prepare once, then overlay the packed cache so paging/persistence
+            // directives reach the loader and the DFlash sidecar stays attached.
+            const prepared = await prepareMuseGlimmerGguf(sourcePath);
+            return this.resolveInternal(prepared, modelType, persistPagedCache);
+        }
         let config;
         try {
             config = JSON.parse(await readFile(join(sourcePath, 'config.json'), 'utf-8'));

package/dist/stream.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"stream.d.ts","sourceRoot":"","sources":["../src/stream.ts"],"names":[],"mappings":"AAEA,OAAO,EACL,WAAW,IAAI,iBAAiB,EAChC,SAAS,IAAI,eAAe,EAC5B,gBAAgB,IAAI,sBAAsB,EAC1C,cAAc,IAAI,oBAAoB,EAEtC,UAAU,IAAI,gBAAgB,EAC9B,WAAW,IAAI,iBAAiB,EAChC,cAAc,IAAI,oBAAoB,EACvC,MAAM,gBAAgB,CAAC;AACxB,OAAO,KAAK,EAIV,eAAe,EACf,gBAAgB,EAChB,kBAAkB,EAElB,cAAc,EACf,MAAM,gBAAgB,CAAC;AAExB,OAAO,KAAK,EAAE,mBAAmB,EAAE,MAAM,mBAAmB,CAAC;AAkC7D,MAAM,WAAW,eAAe;IAC9B,IAAI,EAAE,MAAM,CAAC;IACb,IAAI,EAAE,KAAK,CAAC;IACZ,WAAW,CAAC,EAAE,OAAO,CAAC;CACvB;AAED,MAAM,WAAW,eAAe;IAC9B,IAAI,EAAE,MAAM,CAAC;IACb,IAAI,EAAE,IAAI,CAAC;IACX,YAAY,EAAE,MAAM,CAAC;IACrB,SAAS,EAAE,cAAc,EAAE,CAAC;IAC5B,QAAQ,EAAE,MAAM,GAAG,IAAI,CAAC;IACxB,2EAA2E;IAC3E,eAAe,EAAE,OAAO,CAAC;IACzB,SAAS,EAAE,MAAM,CAAC;IAClB,YAAY,EAAE,MAAM,CAAC;IACrB,eAAe,EAAE,MAAM,CAAC;IACxB,OAAO,EAAE,MAAM,CAAC;IAChB;;;;OAIG;IACH,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB;;;OAGG;IACH,iBAAiB,CAAC,EAAE,OAAO,CAAC;IAC5B;;;;;;;;;;;;;;;;;;;;;OAqBG;IACH,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,WAAW,CAAC,EAAE,kBAAkB,CAAC;CAClC;AAED,MAAM,MAAM,eAAe,GAAG,eAAe,GAAG,eAAe,CAAC;AAKhE,KAAK,oBAAoB,GAAG,eAAe,GAAG,gBAAgB,CAAC;AAE/D,UAAU,qBAAqB;IAC7B,KAAK,EAAE,oBAAoB,CAAC;IAC5B;;;OAGG;IACH,wBAAwB,CAAC,EAAE,MAAM,CAAC;CACnC;AAyDD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAoCG;AACH,wBAAuB,cAAc,CACnC,SAAS,EAAE,CACT,QAAQ,EAAE,CAAC,GAAG,EAAE,KAAK,GAAG,IAAI,EAAE,KAAK,EAAE,eAAe,KAAK,IAAI,KAC1D,OAAO,CAAC,gBAAgB,CAAC,EAC9B,MAAM,CAAC,EAAE,WAAW,GACnB,cAAc,CAAC,eAAe,CAAC,CAsLjC;AAwCD;;;;;;;;;;GAUG;AACH,MAAM,WAAW,uBAAuB;IACtC,sBAAsB,EAAE,CAAC,GAAG,IAAI,EAAE,KAAK,EAAE,KAAK,OAAO,CAAC,gBAAgB,CAAC,CAAC;IACxE,yBAAyB,EAAE,CAAC,GAAG,IAAI,EAAE,KAAK,EAAE,KAAK,OAAO,CAAC,gBAAgB,CAAC,CAAC;IAC3E,6BAA6B,EAAE,CAC7B,GAAG,IAAI,EAAE,KAAK,EAAE,KACb,OAAO,CAAC,gBAAgB,CAAC,CAAC;CAChC;AAED;;;;;;;;;;GAUG;AACH,UAAU,mBAAmB;IAK3B,KAAK,GAAG,IAAI,EAAE,KAAK,EAAE,GAAG,uBAAuB,CAAC;IAKhD,IAAI,CAAC,SAAS,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC,CAAC;IACzC,SAAS,EAAE,uBAAuB,CAAC;CACpC;AAED,mDAAmD;AACnD,UAAU,qBAAqB;IAC7B;;;;;;OAMG;IACH,eAAe,EAAE,OAAO,CAAC;IACzB;;;;;;OAMG;IACH,aAAa,CAAC,EAAE,OAAO,CAAC;IACxB;;;;OAIG;IACH,qBAAqB,CAAC,EAAE,qBAAqB,CAAC;IAC9C;;;;OAIG;IACH,sBAAsB,CAAC,EAAE,OAAO,CAAC;CAClC;AAED;;;;;GAKG;AACH,MAAM,MAAM,cAAc,GAAG,mBAAmB,CAAC;AAEjD;;;;;;GAMG;AACH,KAAK,qBAAqB,CAAC,CAAC,SAAS,qBAAqB,IAAI,CAAC,SAAS;IACtE,aAAa,EAAE,OAAO,CAAC;CACxB,GACG,CAAC,CAAC,eAAe,CAAC,GAClB,CAAC,CAAC,iBAAiB,CAAC,CAAC;AAEzB,sFAAsF;AACtF,MAAM,MAAM,qBAAqB,GAAG,MAAM,uBAAuB,CAAC;AAElE,KAAK,8BAA8B,GAC/B,kBAAkB,GAClB,qBAAqB,GACrB,yBAAyB,GACzB,uBAAuB,GACvB,0BAA0B,GAC1B,8BAA8B,CAAC;AAEnC,KAAK,0BAA0B,CAAC,CAAC,SAAS,qBAAqB,IAC3D,qBAAqB,GACrB,8BAA8B,GAC9B,CAAC,qBAAqB,CAAC,CAAC,CAAC,SAAS,IAAI,GAAG,mBAAmB,GAAG,KAAK,CAAC,CAAC;AAE1E;;;;;;;;;;;;;;GAcG;AACH,MAAM,MAAM,iBAAiB,CAC3B,CAAC,SAAS,mBAAmB,EAC7B,CAAC,SAAS,qBAAqB,IAC7B,IAAI,CAAC,YAAY,CAAC,CAAC,CAAC,EAAE,0BAA0B,CAAC,CAAC,CAAC,CAAC,GACtD,mBAAmB,GACnB,CAAC,qBAAqB,CAAC,CAAC,CAAC,SAAS,IAAI,GAClC,QAAQ,CAAC,IAAI,CAAC,mBAAmB,EAAE,mBAAmB,CAAC,CAAC,GACxD,MAAM,CAAC,CAAC;AAEd;;;;;;;;;;;;;;;;GAgBG;AACH,wBAAgB,kBAAkB,CAChC,CAAC,SAAS,mBAAmB,EAC7B,KAAK,CAAC,CAAC,SAAS,qBAAqB,EAErC,WAAW,EAAE,CAAC,EACd,IAAI,EAAE,CAAC,GACN;IAaD,KAAK,GAAG,IAAI,EAAE,qBAAqB,CAAC,CAAC,CAAC,GAAG,iBAAiB,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;IACjE,IAAI,CAAC,GAAG,IAAI,EAAE,UAAU,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,GAAG,OAAO,CAAC,iBAAiB,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC;CACxE,CAwMA;;;;;;;;;AAED;;;;;;;GAOG;AACH,qBAAa,WAAY,SAAQ,gBAE/B;CAAG;;;;;;;;;AAEL,yEAAyE;AACzE,qBAAa,cAAe,SAAQ,mBAElC;CAAG;;;;;;;;;;;AAEL,8EAA8E;AAC9E,qBAAa,SAAU,SAAQ,cAG7B;CAAG;;;;;;;;;AAEL,0FAA0F;AAC1F,qBAAa,cAAe,SAAQ,mBAElC;CAAG;;;;;;;;;AAEL,gFAAgF;AAChF,qBAAa,WAAY,SAAQ,gBAE/B;CAAG;;;;;;;;;AAEL,yEAAyE;AACzE,qBAAa,gBAAiB,SAAQ,qBAEpC;CAAG;;;;;;;;;;;AAEL;;;;;;GAMG;AACH,qBAAa,UAAW,SAAQ,eAG9B;CAAG"}
+{"version":3,"file":"stream.d.ts","sourceRoot":"","sources":["../src/stream.ts"],"names":[],"mappings":"AAEA,OAAO,EACL,WAAW,IAAI,iBAAiB,EAChC,SAAS,IAAI,eAAe,EAC5B,gBAAgB,IAAI,sBAAsB,EAC1C,cAAc,IAAI,oBAAoB,EAEtC,UAAU,IAAI,gBAAgB,EAC9B,WAAW,IAAI,iBAAiB,EAChC,cAAc,IAAI,oBAAoB,EACvC,MAAM,gBAAgB,CAAC;AACxB,OAAO,KAAK,EAIV,eAAe,EACf,gBAAgB,EAChB,kBAAkB,EAElB,cAAc,EACf,MAAM,gBAAgB,CAAC;AAExB,OAAO,KAAK,EAAE,mBAAmB,EAAE,MAAM,mBAAmB,CAAC;AAkC7D,MAAM,WAAW,eAAe;IAC9B,IAAI,EAAE,MAAM,CAAC;IACb,IAAI,EAAE,KAAK,CAAC;IACZ,WAAW,CAAC,EAAE,OAAO,CAAC;CACvB;AAED,MAAM,WAAW,eAAe;IAC9B,IAAI,EAAE,MAAM,CAAC;IACb,IAAI,EAAE,IAAI,CAAC;IACX,YAAY,EAAE,MAAM,CAAC;IACrB,SAAS,EAAE,cAAc,EAAE,CAAC;IAC5B,QAAQ,EAAE,MAAM,GAAG,IAAI,CAAC;IACxB,2EAA2E;IAC3E,eAAe,EAAE,OAAO,CAAC;IACzB,SAAS,EAAE,MAAM,CAAC;IAClB,YAAY,EAAE,MAAM,CAAC;IACrB,eAAe,EAAE,MAAM,CAAC;IACxB,OAAO,EAAE,MAAM,CAAC;IAChB;;;;OAIG;IACH,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB;;;OAGG;IACH,iBAAiB,CAAC,EAAE,OAAO,CAAC;IAC5B;;;;;;;;;;;;;;;;;;;;;OAqBG;IACH,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,WAAW,CAAC,EAAE,kBAAkB,CAAC;CAClC;AAED,MAAM,MAAM,eAAe,GAAG,eAAe,GAAG,eAAe,CAAC;AAKhE,KAAK,oBAAoB,GAAG,eAAe,GAAG,gBAAgB,CAAC;AAE/D,UAAU,qBAAqB;IAC7B,KAAK,EAAE,oBAAoB,CAAC;IAC5B;;;OAGG;IACH,wBAAwB,CAAC,EAAE,MAAM,CAAC;CACnC;AA+DD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAoCG;AACH,wBAAuB,cAAc,CACnC,SAAS,EAAE,CACT,QAAQ,EAAE,CAAC,GAAG,EAAE,KAAK,GAAG,IAAI,EAAE,KAAK,EAAE,eAAe,KAAK,IAAI,KAC1D,OAAO,CAAC,gBAAgB,CAAC,EAC9B,MAAM,CAAC,EAAE,WAAW,GACnB,cAAc,CAAC,eAAe,CAAC,CAsLjC;AAwCD;;;;;;;;;;GAUG;AACH,MAAM,WAAW,uBAAuB;IACtC,sBAAsB,EAAE,CAAC,GAAG,IAAI,EAAE,KAAK,EAAE,KAAK,OAAO,CAAC,gBAAgB,CAAC,CAAC;IACxE,yBAAyB,EAAE,CAAC,GAAG,IAAI,EAAE,KAAK,EAAE,KAAK,OAAO,CAAC,gBAAgB,CAAC,CAAC;IAC3E,6BAA6B,EAAE,CAC7B,GAAG,IAAI,EAAE,KAAK,EAAE,KACb,OAAO,CAAC,gBAAgB,CAAC,CAAC;CAChC;AAED;;;;;;;;;;GAUG;AACH,UAAU,mBAAmB;IAK3B,KAAK,GAAG,IAAI,EAAE,KAAK,EAAE,GAAG,uBAAuB,CAAC;IAKhD,IAAI,CAAC,SAAS,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC,CAAC;IACzC,SAAS,EAAE,uBAAuB,CAAC;CACpC;AAED,mDAAmD;AACnD,UAAU,qBAAqB;IAC7B;;;;;;OAMG;IACH,eAAe,EAAE,OAAO,CAAC;IACzB;;;;;;OAMG;IACH,aAAa,CAAC,EAAE,OAAO,CAAC;IACxB;;;;OAIG;IACH,qBAAqB,CAAC,EAAE,qBAAqB,CAAC;IAC9C;;;;OAIG;IACH,sBAAsB,CAAC,EAAE,OAAO,CAAC;CAClC;AAED;;;;;GAKG;AACH,MAAM,MAAM,cAAc,GAAG,mBAAmB,CAAC;AAEjD;;;;;;GAMG;AACH,KAAK,qBAAqB,CAAC,CAAC,SAAS,qBAAqB,IAAI,CAAC,SAAS;IACtE,aAAa,EAAE,OAAO,CAAC;CACxB,GACG,CAAC,CAAC,eAAe,CAAC,GAClB,CAAC,CAAC,iBAAiB,CAAC,CAAC;AAEzB,sFAAsF;AACtF,MAAM,MAAM,qBAAqB,GAAG,MAAM,uBAAuB,CAAC;AAElE,KAAK,8BAA8B,GAC/B,kBAAkB,GAClB,qBAAqB,GACrB,yBAAyB,GACzB,uBAAuB,GACvB,0BAA0B,GAC1B,8BAA8B,CAAC;AAEnC,KAAK,0BAA0B,CAAC,CAAC,SAAS,qBAAqB,IAC3D,qBAAqB,GACrB,8BAA8B,GAC9B,CAAC,qBAAqB,CAAC,CAAC,CAAC,SAAS,IAAI,GAAG,mBAAmB,GAAG,KAAK,CAAC,CAAC;AAE1E;;;;;;;;;;;;;;GAcG;AACH,MAAM,MAAM,iBAAiB,CAC3B,CAAC,SAAS,mBAAmB,EAC7B,CAAC,SAAS,qBAAqB,IAC7B,IAAI,CAAC,YAAY,CAAC,CAAC,CAAC,EAAE,0BAA0B,CAAC,CAAC,CAAC,CAAC,GACtD,mBAAmB,GACnB,CAAC,qBAAqB,CAAC,CAAC,CAAC,SAAS,IAAI,GAClC,QAAQ,CAAC,IAAI,CAAC,mBAAmB,EAAE,mBAAmB,CAAC,CAAC,GACxD,MAAM,CAAC,CAAC;AAEd;;;;;;;;;;;;;;;;GAgBG;AACH,wBAAgB,kBAAkB,CAChC,CAAC,SAAS,mBAAmB,EAC7B,KAAK,CAAC,CAAC,SAAS,qBAAqB,EAErC,WAAW,EAAE,CAAC,EACd,IAAI,EAAE,CAAC,GACN;IAaD,KAAK,GAAG,IAAI,EAAE,qBAAqB,CAAC,CAAC,CAAC,GAAG,iBAAiB,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;IACjE,IAAI,CAAC,GAAG,IAAI,EAAE,UAAU,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,GAAG,OAAO,CAAC,iBAAiB,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC;CACxE,CA0MA;;;;;;;;;AAED;;;;;;;GAOG;AACH,qBAAa,WAAY,SAAQ,gBAE/B;CAAG;;;;;;;;;AAEL,yEAAyE;AACzE,qBAAa,cAAe,SAAQ,mBAElC;CAAG;;;;;;;;;;;AAEL,8EAA8E;AAC9E,qBAAa,SAAU,SAAQ,cAG7B;CAAG;;;;;;;;;AAEL,0FAA0F;AAC1F,qBAAa,cAAe,SAAQ,mBAElC;CAAG;;;;;;;;;AAEL,gFAAgF;AAChF,qBAAa,WAAY,SAAQ,gBAE/B;CAAG;;;;;;;;;AAEL,yEAAyE;AACzE,qBAAa,gBAAiB,SAAQ,qBAEpC;CAAG;;;;;;;;;;;AAEL;;;;;;GAMG;AACH,qBAAa,UAAW,SAAQ,eAG9B;CAAG"}

package/dist/stream.js

@@ -5,7 +5,7 @@ const tokenizerPromises = new WeakMap();
 function getNativeIsReasoning(chunk) {
     return typeof chunk.isReasoning === "boolean" ? chunk.isReasoning : undefined;
 }
-async function applyChatTemplateFromModelPath(model, messages, addGenerationPrompt, tools, enableThinking, contentPolicy) {
+async function applyChatTemplateFromModelPath(model, messages, addGenerationPrompt, tools, enableThinking, contentPolicy, reasoningEffort) {
     const modelPath = modelPathsForTokenizers.get(model);
     if (modelPath == null) {
         throw new Error("applyChatTemplate unavailable: model path was not recorded when this model was loaded");
@@ -17,9 +17,9 @@ async function applyChatTemplateFromModelPath(model, messages, addGenerationProm
     }
     const tokenizer = await tokenizerPromise;
     if (contentPolicy == null) {
-        return tokenizer.applyChatTemplate(messages, addGenerationPrompt, tools, enableThinking);
+        return tokenizer.applyChatTemplate(messages, addGenerationPrompt, tools, enableThinking, undefined, undefined, reasoningEffort);
     }
-    return tokenizer.applyChatTemplate(messages, addGenerationPrompt, tools, enableThinking, contentPolicy.order, contentPolicy.existingImagePlaceholder);
+    return tokenizer.applyChatTemplate(messages, addGenerationPrompt, tools, enableThinking, contentPolicy.order, contentPolicy.existingImagePlaceholder, reasoningEffort);
 }
 /**
  * Shared AsyncGenerator adapter for callback-based native streaming methods.
@@ -368,8 +368,8 @@ export function makeStreamingModel(NativeClass, opts) {
         Object.defineProperty(StreamingModelImpl.prototype, "applyChatTemplate", {
             configurable: true,
             writable: true,
-            value(messages, addGenerationPrompt, tools, enableThinking) {
-                return applyChatTemplateFromModelPath(this, messages, addGenerationPrompt, tools, enableThinking, templateContentPolicy);
+            value(messages, addGenerationPrompt, tools, enableThinking, reasoningEffort) {
+                return applyChatTemplateFromModelPath(this, messages, addGenerationPrompt, tools, enableThinking, templateContentPolicy, reasoningEffort);
             },
         });
     }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@mlx-node/lm",
-  "version": "0.0.12",
+  "version": "0.0.15",
   "homepage": "https://github.com/mlx-node/mlx-node",
   "bugs": {
     "url": "https://github.com/mlx-node/mlx-node/issues"
@@ -12,19 +12,37 @@
     "directory": "packages/lm"
   },
   "files": [
-    "dist"
+    "dist",
+    "src"
   ],
   "type": "module",
   "main": "./dist/index.js",
   "types": "./dist/index.d.ts",
   "exports": {
     ".": {
+      "@mlx-node/source": "./src/index.ts",
       "types": "./dist/index.d.ts",
       "import": "./dist/index.js"
     },
+    "./draft-companion": {
+      "@mlx-node/source": "./src/draft-companion.ts",
+      "types": "./dist/draft-companion.d.ts",
+      "import": "./dist/draft-companion.js"
+    },
     "./family-data": {
+      "@mlx-node/source": "./src/family-data.ts",
       "types": "./dist/family-data.d.ts",
       "import": "./dist/family-data.js"
+    },
+    "./model-detection": {
+      "@mlx-node/source": "./src/model-detection.ts",
+      "types": "./dist/model-detection.d.ts",
+      "import": "./dist/model-detection.js"
+    },
+    "./model-discovery": {
+      "@mlx-node/source": "./src/model-discovery.ts",
+      "types": "./dist/model-discovery.d.ts",
+      "import": "./dist/model-discovery.js"
     }
   },
   "scripts": {
@@ -32,7 +50,7 @@
     "test": "vite test run"
   },
   "dependencies": {
-    "@mlx-node/core": "0.0.12"
+    "@mlx-node/core": "0.0.15"
   },
   "devDependencies": {
     "@types/node": "^26.4.0"