@mlx-node/lm 0.0.9 → 0.0.12

package/dist/family-data.js

@@ -0,0 +1,381 @@
+/**
+ * Native-free per-family registration rows plus the pure `matchFamily`
+ * detection they drive.
+ *
+ * This module must stay free of runtime imports (`import type` only): it is
+ * re-exported through the native-free `@mlx-node/agent/catalog` subpath to the
+ * dashboard viewer process, which must never dlopen the Metal addon.
+ * `packages/agent/__test__/catalog-native-free.test.ts` gates that contract.
+ */
+/**
+ * Sampling defaults from Unsloth's Qwen3.6 guide:
+ * https://unsloth.ai/docs/models/qwen3.6#recommended-settings
+ *
+ * All modes pin `top_k = 20` and `min_p = 0.0`; they differ in
+ * `temperature`, `top_p`, and `presence_penalty`.
+ *
+ * Deliberately no `maxConsecutiveTokens` / `maxNgramRepeats` / `ngramSize`: the
+ * native anti-repetition cutoff is off by default (vLLM-aligned), and a client
+ * can still opt in per request.
+ */
+export const QWEN_SAMPLING_DEFAULTS = {
+    /** Thinking mode for precise coding tasks. */
+    thinkingCoding: {
+        temperature: 0.6,
+        topP: 0.95,
+        topK: 20,
+        minP: 0.0,
+        presencePenalty: 0.0,
+        repetitionPenalty: 1.0,
+    },
+    /** Thinking mode for general tasks. */
+    thinkingGeneral: {
+        temperature: 1.0,
+        topP: 0.95,
+        topK: 20,
+        minP: 0.0,
+        presencePenalty: 1.5,
+        repetitionPenalty: 1.0,
+    },
+    /** Instruct (non-thinking) for general tasks. */
+    instructGeneral: {
+        temperature: 0.7,
+        topP: 0.8,
+        topK: 20,
+        minP: 0.0,
+        presencePenalty: 1.5,
+        repetitionPenalty: 1.0,
+    },
+    /** Instruct (non-thinking) for reasoning tasks. */
+    instructReasoning: {
+        temperature: 1.0,
+        topP: 0.95,
+        topK: 20,
+        minP: 0.0,
+        presencePenalty: 1.5,
+        repetitionPenalty: 1.0,
+    },
+};
+/** Sampling defaults for Gemma4 Instruct. */
+export const GEMMA4_SAMPLING_DEFAULTS = {
+    temperature: 0.7,
+    topP: 0.95,
+    topK: 64,
+    minP: 0.0,
+    presencePenalty: 0.0,
+    repetitionPenalty: 1.0,
+};
+/** Sampling defaults from Meta's Muse-Glimmer release recipe. */
+export const MUSE_GLIMMER_SAMPLING_DEFAULTS = {
+    temperature: 0.6,
+    topP: 0.95,
+    topK: 20,
+    minP: 0.0,
+    presencePenalty: 0.0,
+    repetitionPenalty: 1.0,
+};
+/** Sampling defaults from NVIDIA's Nemotron 3.5 Lightning release recipe. */
+export const NEMOTRON_SAMPLING_DEFAULTS = {
+    temperature: 1.0,
+    topP: 0.95,
+    topK: 20,
+    minP: 0.0,
+    presencePenalty: 0.0,
+    repetitionPenalty: 1.0,
+};
+/** Sampling defaults for LFM2.5 Thinking. */
+export const LFM2_SAMPLING_DEFAULTS = {
+    temperature: 0.05,
+    topP: 1.0,
+    topK: 50,
+    minP: 0.0,
+    presencePenalty: 0.0,
+    repetitionPenalty: 1.05,
+};
+/**
+ * Ordered source of truth for every supported model family's registration
+ * data. Each entry owns its canonical `ModelType`, raw config aliases /
+ * architecture probes, and `ChatSession` eligibility via `kind`:
+ *
+ *   - `'trainable'` — GRPO/SFT-capable LM (Qwen3 family); chat-capable.
+ *   - `'loadable'`  — chat-capable LM with no trainer engine (Gemma4, LFM2).
+ *   - `'embedding'` — no chat surface (Harrier); rejected by `loadSession`.
+ *   - `'vlm'`       — VLM whose AsyncGenerator wrapper lives in
+ *                     `@mlx-node/vlm` (importing it here would create a
+ *                     circular package dependency), so `loadSession`
+ *                     rejects it and routes callers to `@mlx-node/vlm`.
+ *
+ * ORDER IS LOAD-BEARING: a base family is selected from an explicit alias or
+ * the single declarative nullish-model_type default, then architecture probes
+ * refine it in declaration order. Gemma's unified architecture is
+ * authoritative (matching the native loader); Harrier refines a Qwen3 base.
+ * Adding a family means adding one data row here plus one loader binding in
+ * `models/model-loader.ts`; the row type and the family-completeness test
+ * enumerate everything else.
+ */
+export const MODEL_FAMILY_DATA = [
+    {
+        id: 'gemma4',
+        kind: 'loadable',
+        match: {
+            rawModelTypes: ['gemma4', 'gemma4_text', 'gemma4_unified'],
+            architectureProbe: ({ architectures }) => architectures.has('Gemma4UnifiedForConditionalGeneration'),
+        },
+        acceptsDraftModel: true,
+        traits: {
+            reasoning: true,
+            thinkingLevelMap: {
+                minimal: 'minimal',
+                low: null,
+                medium: null,
+                high: 'high',
+            },
+            fallbackContextWindow: 131072,
+        },
+        launchPreset: {
+            sampling: GEMMA4_SAMPLING_DEFAULTS,
+            maxOutputTokens: 16384,
+        },
+    },
+    {
+        id: 'muse_glimmer',
+        kind: 'loadable',
+        match: {
+            rawModelTypes: ['muse_glimmer', 'muse_glimmer_text'],
+            architectureProbe: ({ architectures }) => architectures.has('MuseGlimmerForConditionalGeneration'),
+        },
+        traits: {
+            reasoning: true,
+            fallbackContextWindow: 131072,
+        },
+        launchPreset: {
+            sampling: MUSE_GLIMMER_SAMPLING_DEFAULTS,
+            maxOutputTokens: 16384,
+        },
+    },
+    {
+        id: 'harrier',
+        kind: 'embedding',
+        match: {
+            rawModelTypes: ['harrier'],
+            architectureProbe: ({ modelType, architectures }) => modelType === 'qwen3' && architectures.has('Qwen3Model') && !architectures.has('Qwen3ForCausalLM'),
+        },
+    },
+    {
+        id: 'qwen3',
+        kind: 'trainable',
+        match: { rawModelTypes: ['qwen3'] },
+        defaultForNullishModelType: true,
+        traits: { reasoning: true, fallbackContextWindow: 40960 },
+        launchPreset: {
+            sampling: QWEN_SAMPLING_DEFAULTS.thinkingCoding,
+            maxOutputTokens: 38912,
+        },
+    },
+    {
+        id: 'qwen3_5',
+        kind: 'trainable',
+        match: { rawModelTypes: ['qwen3_5'] },
+        acceptsDraftModel: true,
+        ggufArchitectures: ['qwen35'],
+        traits: { reasoning: true, fallbackContextWindow: 262144 },
+        launchPreset: {
+            sampling: QWEN_SAMPLING_DEFAULTS.thinkingCoding,
+            maxOutputTokens: 81920,
+        },
+    },
+    {
+        id: 'qwen3_5_moe',
+        kind: 'trainable',
+        match: { rawModelTypes: ['qwen3_5_moe'] },
+        traits: { reasoning: true, fallbackContextWindow: 262144 },
+        launchPreset: {
+            sampling: QWEN_SAMPLING_DEFAULTS.thinkingCoding,
+            maxOutputTokens: 81920,
+        },
+    },
+    {
+        id: 'lfm2',
+        kind: 'loadable',
+        match: { rawModelTypes: ['lfm2'] },
+        traits: { reasoning: true, fallbackContextWindow: 128000 },
+        launchPreset: {
+            sampling: LFM2_SAMPLING_DEFAULTS,
+            maxOutputTokens: 8192,
+        },
+    },
+    {
+        id: 'lfm2_moe',
+        kind: 'loadable',
+        match: { rawModelTypes: ['lfm2_moe'] },
+        traits: { reasoning: true, fallbackContextWindow: 128000 },
+        /**
+         * LFM2.5-8B-A1B: LiquidAI's MoE card recommends temperature 0.2 / top_k 80
+         * — deliberately NOT the dense `lfm2` values (0.05 / 50).
+         */
+        launchPreset: {
+            sampling: {
+                temperature: 0.2,
+                topP: 1.0,
+                topK: 80,
+                minP: 0.0,
+                presencePenalty: 0.0,
+                repetitionPenalty: 1.05,
+            },
+            maxOutputTokens: 8192,
+        },
+    },
+    {
+        id: 'nemotron_h',
+        kind: 'loadable',
+        match: {
+            rawModelTypes: ['nemotron_h'],
+            architectureProbe: ({ architectures }) => architectures.has('NemotronHForCausalLM'),
+        },
+        traits: {
+            reasoning: true,
+            fallbackContextWindow: 1048576,
+        },
+        launchPreset: {
+            sampling: NEMOTRON_SAMPLING_DEFAULTS,
+            maxOutputTokens: 32768,
+        },
+    },
+    {
+        id: 'internvl_chat',
+        kind: 'vlm',
+        match: { rawModelTypes: ['internvl_chat'] },
+    },
+    {
+        id: 'qianfan-ocr',
+        kind: 'vlm',
+        match: { rawModelTypes: ['qianfan-ocr'] },
+    },
+];
+/**
+ * Every chat-capable family (kind trainable | loadable), in registry order —
+ * the default set the paged-config override manager forces onto the block-paged
+ * path. Derived, so a new chat family can never be forgotten.
+ */
+export const CHAT_FAMILY_IDS = MODEL_FAMILY_DATA.filter((row) => row.kind === 'trainable' || row.kind === 'loadable').map((row) => row.id);
+/** Detection results that cannot back a chat endpoint (kind embedding | vlm). */
+export const NON_GENERATIVE_FAMILY_IDS = new Set(MODEL_FAMILY_DATA.filter((row) => row.kind === 'embedding' || row.kind === 'vlm').map((row) => row.id));
+function buildFamilyDataIndex(rows) {
+    const byId = new Map();
+    const byRawModelType = new Map();
+    let defaultForNullishModelType;
+    for (const family of rows) {
+        const previousFamily = byId.get(family.id);
+        if (previousFamily !== undefined) {
+            throw new Error(`Duplicate canonical model type "${family.id}" in model family registry`);
+        }
+        byId.set(family.id, family);
+        for (const rawModelType of family.match.rawModelTypes) {
+            const previous = byRawModelType.get(rawModelType);
+            if (previous !== undefined) {
+                throw new Error(`Duplicate model_type alias "${rawModelType}" for "${previous.id}" and "${family.id}"`);
+            }
+            byRawModelType.set(rawModelType, family);
+        }
+        if (family.defaultForNullishModelType === true) {
+            if (defaultForNullishModelType !== undefined) {
+                throw new Error(`Duplicate nullish-model_type defaults for "${defaultForNullishModelType.id}" and "${family.id}"`);
+            }
+            defaultForNullishModelType = family;
+        }
+    }
+    if (defaultForNullishModelType === undefined) {
+        throw new Error('Model family registry must declare exactly one nullish-model_type default');
+    }
+    return { byId, byRawModelType, defaultForNullishModelType };
+}
+const FAMILY_DATA_INDEX = buildFamilyDataIndex(MODEL_FAMILY_DATA);
+/** Registration data for a canonical family id, or `undefined` for a foreign string. */
+export function familyDataFor(modelType) {
+    return FAMILY_DATA_INDEX.byId.get(modelType);
+}
+function chatFamilyDataFor(modelType) {
+    const row = FAMILY_DATA_INDEX.byId.get(modelType);
+    return row !== undefined && (row.kind === 'trainable' || row.kind === 'loadable') ? row : undefined;
+}
+/** Canonical family id owning a raw `config.json` model_type alias. */
+export function rawModelTypeToCanonical(rawModelType) {
+    return FAMILY_DATA_INDEX.byRawModelType.get(rawModelType)?.id;
+}
+/** Agent discovery traits for a chat-capable family. */
+export function familyTraitsFor(modelType) {
+    return chatFamilyDataFor(modelType)?.traits;
+}
+/**
+ * Launch preset for every surface that serves a chat family:
+ * `@mlx-node/server` discovery, `mlx launch claude` and `mlx agent`.
+ * `undefined` only for a non-generative family or an unknown type.
+ */
+export function launchPresetFor(modelType) {
+    return chatFamilyDataFor(modelType)?.launchPreset;
+}
+export class MalformedModelConfigError extends Error {
+    constructor(modelPath, reason) {
+        super(`Malformed config.json in ${modelPath}: ${reason}`);
+        this.name = 'MalformedModelConfigError';
+    }
+}
+export class UnsupportedModelTypeError extends Error {
+    constructor(modelPath, rawModelTypeLabel) {
+        super(`Unsupported model_type "${rawModelTypeLabel}" in ${modelPath}/config.json`);
+        this.name = 'UnsupportedModelTypeError';
+    }
+}
+/**
+ * Fail-closed validation: a config.json whose root is not a plain object,
+ * or whose `architectures` is neither an array nor a string, is rejected
+ * instead of coerced (coercion would fall through to the qwen3
+ * nullish-model_type default and silently misroute the checkpoint).
+ */
+function normalizeConfig(modelPath, config) {
+    if (typeof config !== 'object' || config === null || Array.isArray(config)) {
+        throw new MalformedModelConfigError(modelPath, 'root must be a JSON object');
+    }
+    const object = config;
+    const hasModelType = Object.hasOwn(object, 'model_type');
+    const rawModelTypeValue = hasModelType ? object.model_type : undefined;
+    const usesDefaultModelType = !hasModelType || rawModelTypeValue === null;
+    const rawModelType = typeof rawModelTypeValue === 'string' ? rawModelTypeValue : undefined;
+    const rawModelTypeLabel = hasModelType ? String(rawModelTypeValue) : '<missing>';
+    const rawArchitectures = 'architectures' in object ? object.architectures : undefined;
+    if (rawArchitectures !== undefined &&
+        rawArchitectures !== null &&
+        !Array.isArray(rawArchitectures) &&
+        typeof rawArchitectures !== 'string') {
+        throw new MalformedModelConfigError(modelPath, '"architectures" must be an array or a string');
+    }
+    const architectures = Array.isArray(rawArchitectures)
+        ? rawArchitectures.filter((architecture) => typeof architecture === 'string')
+        : typeof rawArchitectures === 'string'
+            ? [rawArchitectures]
+            : [];
+    return { usesDefaultModelType, rawModelType, rawModelTypeLabel, architectures: new Set(architectures) };
+}
+/**
+ * Pure family detection over a parsed `config.json`: alias (or the qwen3
+ * nullish default) picks a base family, then architecture probes refine it in
+ * registry declaration order. Throws {@link MalformedModelConfigError} /
+ * {@link UnsupportedModelTypeError} with `modelPath` naming the checkpoint.
+ * The filesystem/GGUF half lives in `detectModelType`
+ * (`models/model-loader.ts`); native-free consumers (dashboard labels via
+ * `@mlx-node/agent/catalog`) call this directly.
+ */
+export function matchFamily(modelPath, parsedConfig) {
+    const config = normalizeConfig(modelPath, parsedConfig);
+    const rows = MODEL_FAMILY_DATA;
+    const baseFamily = config.usesDefaultModelType
+        ? FAMILY_DATA_INDEX.defaultForNullishModelType
+        : config.rawModelType === undefined
+            ? undefined
+            : FAMILY_DATA_INDEX.byRawModelType.get(config.rawModelType);
+    const matchContext = { ...config, modelType: baseFamily?.id };
+    const family = rows.find((candidate) => candidate.match.architectureProbe?.(matchContext) === true) ?? baseFamily;
+    if (family === undefined)
+        throw new UnsupportedModelTypeError(modelPath, config.rawModelTypeLabel);
+    return family.id;
+}

package/dist/index.d.ts

@@ -14,6 +14,7 @@
  */
 export { Qwen3Model } from './stream.js';
 export { Gemma4Model } from './stream.js';
+export { MuseGlimmerModel } from './stream.js';
 export { HarrierModel } from '@mlx-node/core';
 export { Qwen35Model } from './stream.js';
 export type { Qwen35Config, Qwen35ContextLimits } from '@mlx-node/core';
@@ -21,6 +22,8 @@ export { Lfm2Model } from './stream.js';
 export { LFM2_CONFIGS, getLfm2Config } from './models/lfm2-configs.js';
 export { Qwen35MoeModel } from './stream.js';
 export type { Qwen35MoeConfig } from '@mlx-node/core';
+export { NemotronHModel } from './stream.js';
+export type { NemotronHConfig, NemotronHContextLimits } from '@mlx-node/core';
 export { memoryStats } from '@mlx-node/core';
 export type { ChatConfig, ChatResult, ChatMessage, ToolCallResult, PerformanceMetrics } from '@mlx-node/core';
 export type { ChatStreamDelta, ChatStreamFinal, ChatStreamEvent } from './stream.js';
@@ -29,8 +32,9 @@ export type { NativeStreamingInstance, NativeStreamingMethod, StreamingInstance,
 export { ChatSession, ContextCapacityError, isContextCapacityError } from './chat-session.js';
 export type { ChatSessionOptions, SendOptions, SessionCapableModel, SessionContextLimits } from './chat-session.js';
 export { type Qwen3Config, QWEN3_CONFIGS, type GenerationResult, type GenerationConfig, getQwen3Config, } from './models/qwen3-configs.js';
-export { loadModel, loadSession, detectModelType, type LoadableModel, type TrainableModel, type ModelType, type LoadModelOptions, } from './models/model-loader.js';
-export { AGENT_PAGED_MODEL_TYPES, PagedConfigOverrideManager, QWEN35_PAGED_MODEL_TYPES, type PagedConfigOverrideManagerOptions, } from './models/paged-config-override.js';
+export { loadModel, loadSession, detectModelType, type LoadableModel, type TrainableModel, type LoadModelOptions, } from './models/model-loader.js';
+export { CHAT_FAMILY_IDS, familyDataFor, familyTraitsFor, GEMMA4_SAMPLING_DEFAULTS, launchPresetFor, LFM2_SAMPLING_DEFAULTS, matchFamily, MODEL_FAMILY_DATA, MUSE_GLIMMER_SAMPLING_DEFAULTS, NEMOTRON_SAMPLING_DEFAULTS, NON_GENERATIVE_FAMILY_IDS, QWEN_SAMPLING_DEFAULTS, rawModelTypeToCanonical, type ChatFamilyId, type FamilyThinkingLevelMap, type FamilyTraits, type LaunchPreset, type ModelFamilyData, type ModelFamilyKind, type ModelType, type TrainableFamilyId, } from './family-data.js';
+export { PagedConfigOverrideManager, QWEN35_PAGED_MODEL_TYPES, type PagedConfigOverrideManagerOptions, } from './models/paged-config-override.js';
 export { QWEN35_CONFIGS, getQwen35Config } from './models/qwen3_5-configs.js';
 export * from './tools/index.js';
 export { enableProfiling, disableProfiling } from './profiling.js';

package/dist/index.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../src/index.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;GAaG;AAGH,OAAO,EAAE,UAAU,EAAE,MAAM,aAAa,CAAC;AAGzC,OAAO,EAAE,WAAW,EAAE,MAAM,aAAa,CAAC;AAG1C,OAAO,EAAE,YAAY,EAAE,MAAM,gBAAgB,CAAC;AAC9C,OAAO,EAAE,WAAW,EAAE,MAAM,aAAa,CAAC;AAC1C,YAAY,EAAE,YAAY,EAAE,mBAAmB,EAAE,MAAM,gBAAgB,CAAC;AAGxE,OAAO,EAAE,SAAS,EAAE,MAAM,aAAa,CAAC;AACxC,OAAO,EAAE,YAAY,EAAE,aAAa,EAAE,MAAM,0BAA0B,CAAC;AAGvE,OAAO,EAAE,cAAc,EAAE,MAAM,aAAa,CAAC;AAC7C,YAAY,EAAE,eAAe,EAAE,MAAM,gBAAgB,CAAC;AAsBtD,OAAO,EAAE,WAAW,EAAE,MAAM,gBAAgB,CAAC;AAG7C,YAAY,EAAE,UAAU,EAAE,UAAU,EAAE,WAAW,EAAE,cAAc,EAAE,kBAAkB,EAAE,MAAM,gBAAgB,CAAC;AAG9G,YAAY,EAAE,eAAe,EAAE,eAAe,EAAE,eAAe,EAAE,MAAM,aAAa,CAAC;AASrF,OAAO,EAAE,cAAc,EAAE,kBAAkB,EAAE,MAAM,aAAa,CAAC;AACjE,YAAY,EAAE,uBAAuB,EAAE,qBAAqB,EAAE,iBAAiB,EAAE,cAAc,EAAE,MAAM,aAAa,CAAC;AAMrH,OAAO,EAAE,WAAW,EAAE,oBAAoB,EAAE,sBAAsB,EAAE,MAAM,mBAAmB,CAAC;AAC9F,YAAY,EAAE,kBAAkB,EAAE,WAAW,EAAE,mBAAmB,EAAE,oBAAoB,EAAE,MAAM,mBAAmB,CAAC;AAGpH,OAAO,EACL,KAAK,WAAW,EAChB,aAAa,EACb,KAAK,gBAAgB,EACrB,KAAK,gBAAgB,EACrB,cAAc,GACf,MAAM,2BAA2B,CAAC;AAGnC,OAAO,EACL,SAAS,EACT,WAAW,EACX,eAAe,EACf,KAAK,aAAa,EAClB,KAAK,cAAc,EACnB,KAAK,SAAS,EACd,KAAK,gBAAgB,GACtB,MAAM,0BAA0B,CAAC;AAElC,OAAO,EACL,uBAAuB,EACvB,0BAA0B,EAC1B,wBAAwB,EACxB,KAAK,iCAAiC,GACvC,MAAM,mCAAmC,CAAC;AAE3C,OAAO,EAAE,cAAc,EAAE,eAAe,EAAE,MAAM,6BAA6B,CAAC;AAG9E,cAAc,kBAAkB,CAAC;AAGjC,OAAO,EAAE,eAAe,EAAE,gBAAgB,EAAE,MAAM,gBAAgB,CAAC"}
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../src/index.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;GAaG;AAGH,OAAO,EAAE,UAAU,EAAE,MAAM,aAAa,CAAC;AAGzC,OAAO,EAAE,WAAW,EAAE,MAAM,aAAa,CAAC;AAG1C,OAAO,EAAE,gBAAgB,EAAE,MAAM,aAAa,CAAC;AAG/C,OAAO,EAAE,YAAY,EAAE,MAAM,gBAAgB,CAAC;AAC9C,OAAO,EAAE,WAAW,EAAE,MAAM,aAAa,CAAC;AAC1C,YAAY,EAAE,YAAY,EAAE,mBAAmB,EAAE,MAAM,gBAAgB,CAAC;AAGxE,OAAO,EAAE,SAAS,EAAE,MAAM,aAAa,CAAC;AACxC,OAAO,EAAE,YAAY,EAAE,aAAa,EAAE,MAAM,0BAA0B,CAAC;AAGvE,OAAO,EAAE,cAAc,EAAE,MAAM,aAAa,CAAC;AAC7C,YAAY,EAAE,eAAe,EAAE,MAAM,gBAAgB,CAAC;AAStD,OAAO,EAAE,cAAc,EAAE,MAAM,aAAa,CAAC;AAC7C,YAAY,EAAE,eAAe,EAAE,sBAAsB,EAAE,MAAM,gBAAgB,CAAC;AAsB9E,OAAO,EAAE,WAAW,EAAE,MAAM,gBAAgB,CAAC;AAG7C,YAAY,EAAE,UAAU,EAAE,UAAU,EAAE,WAAW,EAAE,cAAc,EAAE,kBAAkB,EAAE,MAAM,gBAAgB,CAAC;AAG9G,YAAY,EAAE,eAAe,EAAE,eAAe,EAAE,eAAe,EAAE,MAAM,aAAa,CAAC;AASrF,OAAO,EAAE,cAAc,EAAE,kBAAkB,EAAE,MAAM,aAAa,CAAC;AACjE,YAAY,EAAE,uBAAuB,EAAE,qBAAqB,EAAE,iBAAiB,EAAE,cAAc,EAAE,MAAM,aAAa,CAAC;AAMrH,OAAO,EAAE,WAAW,EAAE,oBAAoB,EAAE,sBAAsB,EAAE,MAAM,mBAAmB,CAAC;AAC9F,YAAY,EAAE,kBAAkB,EAAE,WAAW,EAAE,mBAAmB,EAAE,oBAAoB,EAAE,MAAM,mBAAmB,CAAC;AAGpH,OAAO,EACL,KAAK,WAAW,EAChB,aAAa,EACb,KAAK,gBAAgB,EACrB,KAAK,gBAAgB,EACrB,cAAc,GACf,MAAM,2BAA2B,CAAC;AAGnC,OAAO,EACL,SAAS,EACT,WAAW,EACX,eAAe,EACf,KAAK,aAAa,EAClB,KAAK,cAAc,EACnB,KAAK,gBAAgB,GACtB,MAAM,0BAA0B,CAAC;AAKlC,OAAO,EACL,eAAe,EACf,aAAa,EACb,eAAe,EACf,wBAAwB,EACxB,eAAe,EACf,sBAAsB,EACtB,WAAW,EACX,iBAAiB,EACjB,8BAA8B,EAC9B,0BAA0B,EAC1B,yBAAyB,EACzB,sBAAsB,EACtB,uBAAuB,EACvB,KAAK,YAAY,EACjB,KAAK,sBAAsB,EAC3B,KAAK,YAAY,EACjB,KAAK,YAAY,EACjB,KAAK,eAAe,EACpB,KAAK,eAAe,EACpB,KAAK,SAAS,EACd,KAAK,iBAAiB,GACvB,MAAM,kBAAkB,CAAC;AAE1B,OAAO,EACL,0BAA0B,EAC1B,wBAAwB,EACxB,KAAK,iCAAiC,GACvC,MAAM,mCAAmC,CAAC;AAE3C,OAAO,EAAE,cAAc,EAAE,eAAe,EAAE,MAAM,6BAA6B,CAAC;AAG9E,cAAc,kBAAkB,CAAC;AAGjC,OAAO,EAAE,eAAe,EAAE,gBAAgB,EAAE,MAAM,gBAAgB,CAAC"}

package/dist/index.js

@@ -16,6 +16,8 @@
 export { Qwen3Model } from './stream.js';
 // Gemma4 models
 export { Gemma4Model } from './stream.js';
+// Muse-Glimmer models
+export { MuseGlimmerModel } from './stream.js';
 // Embedding models
 export { HarrierModel } from '@mlx-node/core';
 export { Qwen35Model } from './stream.js';
@@ -24,6 +26,14 @@ export { Lfm2Model } from './stream.js';
 export { LFM2_CONFIGS, getLfm2Config } from './models/lfm2-configs.js';
 // MoE variant
 export { Qwen35MoeModel } from './stream.js';
+// Nemotron H models
+//
+// The config/limits types ship alongside the class for the same reason
+// Qwen3.5's do: `NemotronHModel` exposes both `getConfig()` and
+// `contextLimits()`, and without these exports a consumer cannot name
+// either return type. (`MuseGlimmerModel` above deliberately exports no
+// config type — that wrapper has no `getConfig()` at all.)
+export { NemotronHModel } from './stream.js';
 // Memory hygiene: most management is automatic — the decode loop
 // inside `@mlx-node/core` calls `mlx_clear_cache()` every 256 generated
 // tokens to prevent unbounded free-pool growth during long
@@ -64,7 +74,11 @@ export { ChatSession, ContextCapacityError, isContextCapacityError } from './cha
 export { QWEN3_CONFIGS, getQwen3Config, } from './models/qwen3-configs.js';
 // Model loading
 export { loadModel, loadSession, detectModelType, } from './models/model-loader.js';
-export { AGENT_PAGED_MODEL_TYPES, PagedConfigOverrideManager, QWEN35_PAGED_MODEL_TYPES, } from './models/paged-config-override.js';
+// Native-free per-family registration data (also published as the
+// `@mlx-node/lm/family-data` subpath for consumers that must not dlopen the
+// addon; see family-data.ts).
+export { CHAT_FAMILY_IDS, familyDataFor, familyTraitsFor, GEMMA4_SAMPLING_DEFAULTS, launchPresetFor, LFM2_SAMPLING_DEFAULTS, matchFamily, MODEL_FAMILY_DATA, MUSE_GLIMMER_SAMPLING_DEFAULTS, NEMOTRON_SAMPLING_DEFAULTS, NON_GENERATIVE_FAMILY_IDS, QWEN_SAMPLING_DEFAULTS, rawModelTypeToCanonical, } from './family-data.js';
+export { PagedConfigOverrideManager, QWEN35_PAGED_MODEL_TYPES, } from './models/paged-config-override.js';
 export { QWEN35_CONFIGS, getQwen35Config } from './models/qwen3_5-configs.js';
 // Tool calling utilities
 export * from './tools/index.js';