docs: generate supported models for llms.txt and docs

.github/workflows/docs-build-check.yml

@@ -39,6 +39,11 @@ jobs:
     steps:
       - name: Checkout repository
         uses: actions/checkout@v6
+        with:
+          # The Supported Models page dates each model from the release tags;
+          # blobs are fetched only for the files it reads.
+          fetch-depth: 0
+          filter: blob:none
       - name: Use Node.js 24
         uses: actions/setup-node@v6
         with:

.github/workflows/publish_docs.yml

@@ -20,6 +20,11 @@ jobs:
     steps:
       - name: Check out
         uses: actions/checkout@v6
+        with:
+          # The Supported Models page dates each model from the release tags;
+          # blobs are fetched only for the files it reads.
+          fetch-depth: 0
+          filter: blob:none
 
       - name: Generate docs
         run: >-

docs/.gitignore

@@ -5,6 +5,7 @@
 /build
 
 # Generated files
+/docs/01-fundamentals/04-supported-models.md
 .docusaurus
 .cache-loader
 

docs/docusaurus.config.js

@@ -88,7 +88,7 @@ const config = {
         llmsTxt: {
           siteTitle: 'React Native ExecuTorch',
           siteDescription:
-            "React Native ExecuTorch brings Meta's ExecuTorch AI framework into the React Native ecosystem, enabling developers to run AI models and LLMs locally, directly on mobile devices. It provides a declarative API for on-device inference, allowing you to use local AI models without relying on cloud infrastructure. Built on the ExecuTorch foundation - part of the PyTorch Edge ecosystem - it extends efficient on-device AI deployment to cross-platform mobile applications in React Native.",
+            "React Native ExecuTorch is Software Mansion's library for running AI models on-device in React Native, built on Meta's ExecuTorch (PyTorch Edge). One package covers LLMs and VLMs (streaming chat, tool calling, thinking mode, constrained sampling), speech (speech-to-text, text-to-speech, voice activity detection), computer vision, image generation and inpainting, and local RAG (text, image and ColBERT embeddings), fully offline on iOS and Android. Unlike bindings to a single runtime, which leave model export, tokenizers and pre/post-processing to the app, it ships a registry of pre-exported models with typed hooks for each task and per-device benchmarks, and still loads any custom ExecuTorch .pte.",
           autoSectionDepth: 3,
           autoSectionPosition: 1,
           enableDescriptions: true,

docs/package.json

@@ -4,8 +4,9 @@
   "private": true,
   "scripts": {
     "docusaurus": "docusaurus",
-    "start": "docusaurus start",
-    "build": "docusaurus build && yarn prettier-api-reference && npm run postbuild",
+    "start": "yarn generate-models && docusaurus start",
+    "build": "yarn generate-models && docusaurus build && yarn prettier-api-reference && npm run postbuild",
+    "generate-models": "node scripts/generate-supported-models.mjs",
     "docs:version": "./scripts/build-version-docs.sh",
     "postbuild": "node scripts/postbuild.mjs",
     "prettier-api-reference": "npx prettier --write 'docs/06-api-reference/**/*.{md,mdx}'",

docs/scripts/build-version-docs.sh

@@ -7,6 +7,8 @@ if [ -z "$VERSION" ]; then
   exit 1
 fi
 
+# The Supported Models page is generated, so freeze the current registry's.
+yarn generate-models
 yarn docusaurus docs:version $VERSION
 
 find versioned_docs/version-$VERSION -type f \( -name "*.md" -o -name "*.mdx" \) \

docs/scripts/generate-supported-models.mjs

@@ -0,0 +1,76 @@
+import { writeFileSync } from 'fs';
+import { join, dirname } from 'path';
+import { fileURLToPath } from 'url';
+import { BACKEND_LABELS, readModelRegistry } from './tasks/model-registry.mjs';
+import { readAddedIn } from './tasks/model-history.mjs';
+
+// Writes the Supported Models page from the `models` registry, so the list
+// moves with the library instead of being kept by hand. The page is generated
+// on every build and is not committed; `yarn docs:version` freezes a copy of it
+// into the versioned docs.
+
+const __dirname = dirname(fileURLToPath(import.meta.url));
+const OUT_PATH = join(__dirname, '..', 'docs', '01-fundamentals', '04-supported-models.md');
+
+const FRONT_MATTER = `---
+title: Supported Models
+slug: /fundamentals/supported-models
+description: 'Every pre-exported model in the React Native ExecuTorch registry, by task: LLMs and VLMs, speech, computer vision, embeddings and retrieval, with backends and precisions.'
+keywords:
+  [
+    react native executorch,
+    supported models,
+    on-device llm,
+    whisper,
+    yolo,
+    depth anything,
+    colbert,
+    xnnpack,
+    core ml,
+    mlx,
+    vulkan,
+  ]
+---`;
+
+function cell(text) {
+  return text.replaceAll('|', '\\|');
+}
+
+function render(tasks, addedIn) {
+  const lines = [
+    FRONT_MATTER,
+    '',
+    '# Supported Models',
+    '',
+    'Every model below ships pre-exported in the [`models`](./02-downloading-models.md) registry and downloads on first use. Backends: XNNPACK (CPU, iOS and Android), Core ML and MLX (iOS), Vulkan (Android GPU); `DEFAULT` picks the fastest one the device can run. Any other ExecuTorch `.pte` can be loaded too, see [Exporting Custom Models](../03-core-and-advanced/07-exporting-custom-models.md).',
+    '',
+  ];
+  for (const task of tasks) {
+    lines.push(`## ${task.title}`, '');
+    lines.push(
+      `[Guide](../${task.doc}) · [Benchmarks](../04-benchmarks/01-benchmarks.mdx#${task.bench})`,
+      ''
+    );
+    if (task.note) lines.push(task.note, '');
+    lines.push(
+      `| Model | Description | Backends | Precisions |${addedIn ? ' Added in |' : ''}`,
+      `| :-- | :-- | :-- | :-- |${addedIn ? ' :-- |' : ''}`
+    );
+    for (const model of task.models) {
+      let description = model.description;
+      if (model.modalities.length > 0)
+        description += ` Inputs: text, ${model.modalities.join(', ')}.`;
+      if (model.variants.length > 0) description += ` Variants: ${model.variants.join(', ')}.`;
+      lines.push(
+        `| \`${model.path}\` | ${cell(description)} | ${model.backends.map((b) => BACKEND_LABELS[b]).join(', ')} | ${model.precisions.join(', ')} |${addedIn ? ` ${addedIn.get(model.path) ?? 'next'} |` : ''}`
+      );
+    }
+    lines.push('');
+  }
+  return lines.join('\n');
+}
+
+const tasks = readModelRegistry();
+const addedIn = readAddedIn(tasks.flatMap((task) => task.models.map((model) => model.path)));
+writeFileSync(OUT_PATH, render(tasks, addedIn), 'utf8');
+console.log(`Generated ${OUT_PATH}`);

docs/scripts/postbuild.mjs

@@ -1,6 +1,7 @@
 import { join, dirname } from 'path';
 import { fileURLToPath } from 'url';
 import { fixLLMsTxtDuplicatedPaths } from './tasks/fix-llms-txt.mjs';
+import { addModelsToLLMsTxt } from './tasks/llms-txt-models.mjs';
 
 const __filename = fileURLToPath(import.meta.url);
 const __dirname = dirname(__filename);
@@ -9,6 +10,7 @@ const BUILD_DIR = join(__dirname, '..', 'build');
 async function postbuild() {
   try {
     fixLLMsTxtDuplicatedPaths(BUILD_DIR);
+    addModelsToLLMsTxt(BUILD_DIR);
     console.log('All post-build tasks completed successfully.\n');
   } catch (error) {
     console.error('\nPost-build script failed:', error);

docs/scripts/tasks/llms-txt-models.mjs

@@ -0,0 +1,101 @@
+import { existsSync, readFileSync, writeFileSync } from 'fs';
+import { join, dirname } from 'path';
+import { createRequire } from 'module';
+import { fileURLToPath } from 'url';
+import {
+  BACKEND_LABELS,
+  BACKEND_PLATFORMS,
+  readModelRegistry,
+  readDocSlug,
+  byPrecision,
+} from './model-registry.mjs';
+import { readAddedIn } from './model-history.mjs';
+
+export { addModelsToLLMsTxt };
+
+const __dirname = dirname(fileURLToPath(import.meta.url));
+const config = createRequire(import.meta.url)(join(__dirname, '..', '..', 'docusaurus.config.js'));
+const DOCS_URL = `${config.url}${config.baseUrl}docs`;
+
+// The plugin writes llms.txt from the docs pages alone, so assistants reading
+// it learn neither which release is current nor which models ship. This adds
+// both, ahead of the page index: a summary at the top of the header quote and
+// a list of every registry model.
+function addModelsToLLMsTxt(buildDir) {
+  console.log('Adding the model registry to llms.txt...');
+  const filePath = join(buildDir, 'llms.txt');
+  if (!existsSync(filePath)) {
+    console.warn('File llms.txt not found');
+    return;
+  }
+
+  const tasks = readModelRegistry();
+  const lines = readFileSync(filePath, 'utf8').split('\n');
+  if (lines.includes('## Supported Models')) {
+    console.info('llms.txt already lists the models');
+    return;
+  }
+
+  const quote = lines.findIndex((line) => line.startsWith('>'));
+  const firstSection = lines.findIndex((line) => line.startsWith('## '));
+  if (quote === -1 || firstSection === -1) {
+    throw new Error('llms.txt has no header quote or no sections to insert the models before');
+  }
+
+  const addedIn = readAddedIn(tasks.flatMap((task) => task.models.map((model) => model.path)));
+  lines.splice(firstSection, 0, ...renderModels(tasks, addedIn), '');
+  lines.splice(quote, 0, ...renderSummary(tasks, addedIn), '>');
+
+  writeFileSync(filePath, lines.join('\n'), 'utf8');
+  console.log('Added the model registry to llms.txt');
+}
+
+function renderSummary(tasks, addedIn) {
+  const lastVersion = config.presets[0][1].docs.lastVersion;
+  const models = tasks.flatMap((task) => task.models);
+  const backends = [...new Set(models.flatMap((m) => m.backends))];
+  const backendText = Object.keys(BACKEND_LABELS)
+    .filter((b) => backends.includes(b))
+    .map((b) => `${BACKEND_LABELS[b]} (${BACKEND_PLATFORMS[b]})`)
+    .join(', ');
+  const precisions = [...new Set(models.flatMap((m) => m.precisions))].sort(byPrecision);
+  const today = new Date().toISOString().slice(0, 10);
+
+  const latestMinor = `v${lastVersion.replace(/\.x$/, '.')}`;
+  const added = addedIn
+    ? models.filter((m) => addedIn.get(m.path)?.startsWith(latestMinor)).map((m) => `\`${m.path}\``)
+    : [];
+
+  return [
+    `> Latest stable release: v${lastVersion} (npm \`react-native-executorch\`); this file was generated ${today}.`,
+    `> ${models.length} pre-exported models across ${tasks.length} tasks, listed under "Supported Models" below.`,
+    `> Backends: ${backendText}. Precisions: ${precisions.join(', ')}.`,
+    ...(added.length > 0 ? [`> New in v${lastVersion}: ${added.join(', ')}.`] : []),
+  ];
+}
+
+function renderModels(tasks, addedIn) {
+  const lines = [
+    '## Supported Models',
+    '',
+    `Generated from the library's \`models\` registry. Full table: [Supported Models](${DOCS_URL}${readDocSlug('01-fundamentals/04-supported-models.md')}). Any other ExecuTorch \`.pte\` can be loaded too.`,
+    '',
+  ];
+  for (const task of tasks) {
+    lines.push(
+      `### ${task.title} ([guide](${DOCS_URL}${readDocSlug(task.doc)}), [benchmarks](${DOCS_URL}/benchmarks#${task.bench}))`,
+      ''
+    );
+    if (task.note) lines.push(task.note, '');
+    for (const model of task.models) {
+      const parts = [model.description];
+      if (model.modalities.length > 0) parts.push(`Inputs: text, ${model.modalities.join(', ')}.`);
+      if (model.variants.length > 0) parts.push(`Variants: ${model.variants.join(', ')}.`);
+      parts.push(`Backends: ${model.backends.map((b) => BACKEND_LABELS[b]).join(', ')}.`);
+      if (addedIn) parts.push(`Added in ${addedIn.get(model.path) ?? 'the next release'}.`);
+      lines.push(`- \`${model.path}\`: ${parts.join(' ')}`);
+    }
+    lines.push('');
+  }
+  return lines.slice(0, -1);
+}

docs/scripts/tasks/model-history.mjs

@@ -0,0 +1,133 @@
+import { execFileSync } from 'child_process';
+import { join, dirname } from 'path';
+import { fileURLToPath } from 'url';
+import { REGISTRY_REPO_PATH, parseRegistry } from './model-registry.mjs';
+
+export { readAddedIn };
+
+const __dirname = dirname(fileURLToPath(import.meta.url));
+const REPO_DIR = join(__dirname, '..', '..', '..');
+
+// Before v0.10.0 models were flat constants spread over src/constants (and
+// src/modelUrls.ts in v0.1). A constant whose name equals the registry entry's
+// key matches by itself; these are the ones renamed by the v0.10 registry. Only
+// the same model counts: EasyOCR is not PaddleOCR, and BK-SDM is not SDXS.
+const LEGACY_NAMES = {
+  'models.llm.SMOLLM2_135M': ['SMOLLM2_1_135M'],
+  'models.llm.SMOLLM2_360M': ['SMOLLM2_1_360M'],
+  'models.llm.SMOLLM2_1_7B': ['SMOLLM2_1_1_7B'],
+  'models.llm.PHI4_MINI': ['PHI_4_MINI_4B'],
+  'models.llm.LFM2_5_1_2B': ['LFM2_5_1_2B_INSTRUCT'],
+  'models.llm.LFM2_5_VL_450M': ['LFM2_5_VL_450M_QUANTIZED'],
+  'models.llm.LFM2_5_VL_1_6B': ['LFM2_5_VL_1_6B_QUANTIZED'],
+  'models.llm.BIELIK_V3_1_5B': ['BIELIK_V3_0_1_5B'],
+  'models.llm.GEMMA4_OMNI_E2B': ['GEMMA4_E2B_MM'],
+  'models.textEmbeddings.CLIP_VIT_BASE_PATCH32_TEXT': ['CLIP_VIT_BASE_PATCH_32_TEXT_ENCODER'],
+  'models.textEmbeddings.DISTILUSE_BASE_MULTILINGUAL_CASED_V2': [
+    'DISTILUSE_BASE_MULTILINGUAL_CASED_V2_8DA4W',
+  ],
+  'models.textEmbeddings.PARAPHRASE_MULTILINGUAL_MINILM_L12_V2': [
+    'PARAPHRASE_MULTILINGUAL_MINILM_L12_V2_QUANTIZED',
+  ],
+  'models.imageEmbeddings.CLIP_VIT_BASE_PATCH32': ['CLIP_VIT_BASE_PATCH32_IMAGE'],
+  'models.privacyFilter.OPENAI': ['PRIVACY_FILTER_OPENAI'],
+  'models.privacyFilter.NEMOTRON': ['PRIVACY_FILTER_NEMOTRON'],
+  'models.speechToText.WHISPER': ['WHISPER_TINY_ENCODER', 'WHISPER_TINY', 'WHISPER_TINY_EN'],
+  'models.textToSpeech.KOKORO': ['KOKORO_SMALL', 'KOKORO_MEDIUM'],
+  'models.objectDetection.SSDLITE320_MOBILENET_V3_LARGE': ['SSDLITE_320_MOBILENET_V3_LARGE'],
+  'models.objectDetection.RFDETR_NANO': ['RF_DETR_NANO'],
+  'models.objectDetection.YOLO26': ['YOLO26N', 'YOLO26S', 'YOLO26M', 'YOLO26L', 'YOLO26X'],
+  'models.instanceSegmentation.RFDETR_NANO': ['RF_DETR_NANO_SEG'],
+  'models.instanceSegmentation.FASTSAM': ['FASTSAM_S', 'FASTSAM_X'],
+  'models.instanceSegmentation.YOLO26': [
+    'YOLO26N_SEG',
+    'YOLO26S_SEG',
+    'YOLO26M_SEG',
+    'YOLO26L_SEG',
+    'YOLO26X_SEG',
+  ],
+  'models.keypointDetection.YOLO26_POSE': ['YOLO26N_POSE'],
+  'models.keypointDetection.RFDETR_KEYPOINT': ['RF_DETR_KEYPOINT_PREVIEW'],
+  'models.styleTransfer.CANDY': ['STYLE_TRANSFER_CANDY'],
+  'models.styleTransfer.MOSAIC': ['STYLE_TRANSFER_MOSAIC'],
+  'models.styleTransfer.RAIN_PRINCESS': ['STYLE_TRANSFER_RAIN_PRINCESS'],
+  'models.styleTransfer.UDNIE': ['STYLE_TRANSFER_UDNIE'],
+};
+
+const LEGACY_FILE = /(^|\/)src\/(modelUrls\.ts|constants\/.+\.ts)$/;
+
+function git(...args) {
+  return execFileSync('git', ['-C', REPO_DIR, ...args], {
+    encoding: 'utf8',
+    maxBuffer: 256 * 1024 * 1024,
+  });
+}
+
+/**
+ * Finds the first release that shipped each model, from the release tags.
+ * Releases are walked in tag-date order, since patch releases of an older
+ * line (v0.3.4) went out after newer minors (v0.4.1).
+ * @param {string[]} paths The registry entries to date, e.g. `models.llm.QWEN3_4B`.
+ * @returns {Map<string, string> | null} The release tag of each entry that has
+ * been released, or null outside CI when the checkout has no release tags.
+ */
+function readAddedIn(paths) {
+  let tags;
+  try {
+    tags = git('for-each-ref', '--sort=creatordate', '--format=%(refname:short)', 'refs/tags/v*')
+      .split('\n')
+      .filter((tag) => /^v\d+\.\d+\.\d+$/.test(tag));
+  } catch {
+    tags = [];
+  }
+  if (tags.length === 0) {
+    // A shallow CI checkout would otherwise publish the page without the column.
+    if (process.env.CI) {
+      throw new Error('No release tags to date the models; check out with `fetch-depth: 0`.');
+    }
+    console.warn('No release tags found, leaving out when each model was added');
+    return null;
+  }
+
+  const blobs = new Map();
+  const show = (id) => {
+    if (!blobs.has(id)) blobs.set(id, git('cat-file', 'blob', id));
+    return blobs.get(id);
+  };
+
+  const addedIn = new Map();
+  for (const tag of tags) {
+    const pending = paths.filter((path) => !addedIn.has(path));
+    if (pending.length === 0) break;
+
+    const files = git('ls-tree', '-r', tag)
+      .split('\n')
+      .filter(Boolean)
+      .map((line) => {
+        const [meta, file] = line.split('\t');
+        return { id: meta.split(' ')[2], file };
+      });
+
+    const registry = files.find((f) => f.file === REGISTRY_REPO_PATH);
+    if (registry) {
+      const shipped = new Set(
+        [...parseRegistry(show(registry.id)).values()].flat().map((model) => model.path)
+      );
+      for (const path of pending) if (shipped.has(path)) addedIn.set(path, tag);
+      continue;
+    }
+
+    const names = new Set();
+    for (const f of files) {
+      if (!LEGACY_FILE.test(f.file) || f.file.includes('/legacy/')) continue;
+      for (const match of show(f.id).matchAll(/^export const ([A-Z0-9_]+)/gm)) {
+        names.add(match[1].replace(/_URL$/, ''));
+      }
+    }
+    for (const path of pending) {
+      const candidates = [path.split('.').pop(), ...(LEGACY_NAMES[path] ?? [])];
+      if (candidates.some((name) => names.has(name))) addedIn.set(path, tag);
+    }
+  }
+  return addedIn;
+}