NakliTechie commited on
Commit
869ba2d
·
1 Parent(s): b0d59cd

Sync from GitHub 2026-10-30T05:02:25Z

Browse files
Files changed (1) hide show
  1. index.html +168 -27
index.html CHANGED
@@ -866,6 +866,7 @@
866
  height: 64px;
867
  }
868
 
 
869
  .attachment-chip.audio-chip {
870
  width: 180px;
871
  height: 48px;
@@ -2501,7 +2502,7 @@
2501
  <li><strong>Artifact preview</strong> &mdash; <code>```html</code> / <code>```svg</code> / <code>```artifact</code> code blocks get a live sandboxed iframe below the code (<code>sandbox=&quot;allow-scripts&quot;</code>, no same-origin). Safe to run model-generated UI inline.</li>
2502
  <li><strong>Voice to text</strong> &mdash; 🗣 button left of the input records mic audio, decodes to 16 kHz mono PCM on-device, and runs Whisper-base via WebGPU to transcribe into the input. ~80 MB first-use download; nothing leaves the device.</li>
2503
  <li><strong>Read aloud</strong> &mdash; 🔊 button under any reply speaks it via Kokoro-82M (runs on WASM, so it never competes with the chat model). Streams sentence-by-sentence; English voices in Settings, and a Hindi reply uses the device&rsquo;s own Hindi voice when it has one. ~80 MB first-use download; nothing leaves the device.</li>
2504
- <li><strong>Voice mode</strong> &mdash; 🎙 chip opens a hands-free loop: hold the mic (or hold <strong>V</strong>) to talk, release to send. Moonshine transcribes on-device (English; other languages fall back to Whisper), the model answers, and Kokoro reads the reply back. Turn-based; the chat model stays loaded. ~180 MB first-use download.</li>
2505
  <li><strong>Python code tool</strong> (agent-capable models) &mdash; model can call <code>run_python</code> to execute Python in a sandboxed Pyodide worker. numpy / pandas / matplotlib auto-install on import. ~10 MB first-use download.</li>
2506
  <li><strong>Cache management</strong> &mdash; view/clear cached models in Settings</li>
2507
  <li><strong>Custom models</strong> &mdash; Settings &rarr; paste a Hugging Face ONNX repo id (causal LMs, runs on WebGPU) <em>or a direct GGUF URL</em> (runs via wllama, WebGPU when present / CPU otherwise). ONNX repos are probed via the HF API for the best quant; oversized models warn (but aren&rsquo;t blocked) so you can try on a capable GPU. GGUF URLs are sized via a HEAD request; point at the first shard of a split GGUF (&hellip;-00001-of-&hellip;) for files over wllama&rsquo;s 2 GB per-file / ~4 GB total limits.</li>
@@ -2680,7 +2681,7 @@
2680
  <li><strong>Gemma 4 E2B &middot; WebGPU kernels</strong> (~2 GB) &mdash; text + agent. Google's Gemma 4 E2B (QAT mobile, int4) on a from-scratch WebGPU engine: hand-written WGSL for QAT int4 matmul, embed-gather-norm, RoPE/RMSNorm and GQA + sliding-window attention, reading the QAT-mobile weights directly. Ported from <em>webml-community/gemma-4-webgpu-kernels</em>. WebGPU-only. About 4&times; the ONNX Gemma 4 E2B&rsquo;s decode speed on the same Mac (171 vs 42 tok/s, M4 Pro); text-only.</li>
2681
  <li><strong>Gemma 4 E4B &middot; WebGPU kernels</strong> (~3.6 GB) &mdash; text + agent. The larger Gemma 4 (QAT mobile) on the same hand-written WebGPU engine. Its per-layer embedding table can also live on disk (Settings &rarr; Models, below).</li>
2682
  <li><strong>LFM3-VL 2B</strong> (~1.9 GB) &mdash; text and images, with reasoning. Liquid AI's LFM3-VL-2B (October 2026): the LFM3 language model with a new vision encoder, through Transformers.js on WebGPU. Attach an image and ask about it. Show reasoning turns its step-by-step thinking on.</li>
2683
- <li><strong>LFM3-Omni 2B</strong> (~2.1 GB) &mdash; text, images and speech in, with reasoning and tools. Liquid AI's LFM3-Omni-2B (October 2026), one model that reads text, pictures and spoken questions (up to 30 seconds a clip). Transformers.js does not run it yet, so it comes through anyformer on diskformer's own onnxruntime-web engine; the first load stores the files in this browser once. It answers in text for now; spoken answers come later.</li>
2684
  <li><strong>Gemma 4 E2B</strong> (~1.5 GB) &mdash; multimodal (image + audio) + agent.</li>
2685
  <li><strong>Gemma 4 E4B</strong> (~4.9 GB) &mdash; multimodal + agent, best quality.</li>
2686
  <li><strong>LFM2.5 230M &middot; GGUF</strong> (~170 MB) &mdash; the same LFM2.5 230M through wllama (llama.cpp &rarr; WebAssembly), reading a Q5_K_M GGUF. WebGPU when your browser has it, CPU otherwise, so it also runs where WebGPU is missing.</li>
@@ -2850,6 +2851,10 @@
2850
  <input type="checkbox" id="toolApprovalToggle" checked>
2851
  Ask before running custom and MCP tools <span style="font-weight:400;color:var(--gray-500)">(they send the model's arguments to another server; also asks before weather and Wikipedia lookups)</span>
2852
  </label>
 
 
 
 
2853
  <label class="toggle-row" id="keylessWebRow">
2854
  <input type="checkbox" id="keylessWebToggle" checked>
2855
  Keyless web tools: weather and Wikipedia <span style="font-weight:400;color:var(--gray-500)">(agent-capable models; sends the place name to open-meteo.com or the topic to wikipedia.org, no key or account)</span>
@@ -3005,7 +3010,7 @@
3005
  <option value="lfm3-1.7b">LFM3 1.7B &middot; Agent (~1.4 GB)</option>
3006
  <option value="gemini-nano">Gemini Nano &middot; Built into Chrome (0 MB)</option>
3007
  <option value="lfm3-vl-2b">LFM3-VL 2B &middot; Text and images (~1.9 GB)</option>
3008
- <option value="lfm3-omni-2b-anyformer">LFM3-Omni 2B &middot; Text, images and speech in (~2.1 GB)</option>
3009
  <option value="gemma4-e2b">Gemma 4 E2B &middot; Multimodal (~1.5 GB)</option>
3010
  <option value="gemma4-e4b">Gemma 4 E4B &middot; Multimodal (~4.9 GB)</option>
3011
  <option value="lfm2-230m-gguf">LFM2.5 230M &middot; GGUF (~170 MB)</option>
@@ -3170,7 +3175,11 @@
3170
  </div>
3171
  </div>
3172
  <div class="mode-extra" id="voicePanel" role="region" aria-label="Voice conversation">
3173
- <div class="mode-extra-head">Voice &middot; <span class="memory-count">Moonshine STT + Kokoro TTS &middot; on-device</span></div>
 
 
 
 
3174
  <div class="voice-stage">
3175
  <button type="button" class="voice-mic" id="voiceMicBtn" aria-label="Hold to talk, release to send">
3176
  <span class="voice-mic-icon" id="voiceMicIcon">&#127897;&#65039;</span>
@@ -8031,7 +8040,7 @@ const engineFetch = async (url, init = {}) => {
8031
  // The model loads through anyformer (convergence plan, 2026-10-22): anyformer resolves the repo to diskformer's copy
8032
  // of the engine (Gemma4Mobile) and, with ple, builds the OPFS per-layer-embedding table itself. The worker then drives
8033
  // session.engine exactly as it drove its own engine. Pinned to the same anyformer as anyformer_models.js.
8034
- const ANYFORMER_URL = 'https://cdn.jsdelivr.net/npm/anyformer@0.4.13/index.js';
8035
 
8036
  let anyformerLoad = null;
8037
  let session = null;
@@ -8176,7 +8185,7 @@ self.onmessage = async (e) => {
8176
  // rolled-in module URLs (moduleUrl()).
8177
  // The model loads through anyformer (convergence plan, 2026-10-22): anyformer resolves the repo to diskformer's copy of
8178
  // the engine (TernaryBonsai2); the worker drives session.engine as before and attaches DFlash / MTP itself.
8179
- const ANYFORMER_URL = 'https://cdn.jsdelivr.net/npm/anyformer@0.4.13/index.js';
8180
  const DFLASH_URL = '__TERNARY_BONSAI2_DFLASH_URL__';
8181
  const MTP_URL = '__TERNARY_BONSAI2_MTP_URL__';
8182
  // MTP speculative decoding (Settings → Models, default OFF; posted as `mtp` + `mtpDepth` on the
@@ -8699,6 +8708,12 @@ self.onmessage = async (e) => {
8699
  try { model.reset(); } catch (_) {}
8700
  }
8701
  lastHistoryLen = msgs.length;
 
 
 
 
 
 
8702
  abortController = new AbortController();
8703
  const open = model.thinkOpenTokenId, close = model.thinkCloseTokenId;
8704
  // A reasoning-only model (LFM2.5 24B: alwaysThinks) opens <think> in every prompt, so its reply always starts as reasoning.
@@ -8710,11 +8725,18 @@ self.onmessage = async (e) => {
8710
  maxNewTokens: cfg.max_new_tokens || cfg.max_tokens || 1024,
8711
  signal: abortController.signal,
8712
  enableThinking: wantThink,
 
8713
  });
 
8714
  for await (const out of stream) {
8715
  if (stopFlag) break;
8716
  const full = out && typeof out.text === 'string' ? out.text : '';
8717
  last = full;
 
 
 
 
 
8718
  // A multi-byte character split across tokens decodes as a trailing U+FFFD until its last byte
8719
  // arrives. Hold the delta back until then: posting it sent '\uFFFD' and then cut the next delta
8720
  // one code unit late, so 🍂 arrived as '\uFFFD\uDF42' (2026-10-14).
@@ -9488,7 +9510,7 @@ self.addEventListener('message', async (e) => {
9488
  id: 'LiquidAI/LFM3-Omni-2B-ONNX',
9489
  label: 'LFM3-Omni 2B',
9490
  backend: 'anyformer-lfm3-omni',
9491
- size: '~2.1 GB',
9492
  type: 'causal',
9493
  multimodal: true,
9494
  imageInput: true,
@@ -19035,6 +19057,10 @@ self.addEventListener('message', async (e) => {
19035
  }
19036
  }
19037
 
 
 
 
 
19038
  else if (data.type === 'token') {
19039
  // Always accumulate so non-streaming callers (chat UI, non-streaming
19040
  // API) still receive the full text via the 'complete' handler.
@@ -19847,6 +19873,7 @@ self.addEventListener('message', async (e) => {
19847
  attachments: attData || null,
19848
  enableThinking,
19849
  generationConfig: genConfig,
 
19850
  });
19851
  });
19852
  }
@@ -20143,10 +20170,14 @@ self.addEventListener('message', async (e) => {
20143
  // a flag for whether any tool was actually called.
20144
  // TOOL-APPROVAL (Open Q#11): custom and MCP tools send the model's arguments to another server, so each call
20145
  // waits for Allow / Always / Deny on a card in the chat, unless the gate is off (Settings → Tools) or the user
20146
- // chose "Always" for that tool this session. The keyless web built-ins (get_weather, wikipedia_lookup) carry an
20147
- // approvalHost and ask the same way; every other built-in runs without a card.
20148
- const TOOL_APPROVAL_KEY = 'lm_tool_approval';
20149
- const toolApprovedThisSession = new Set();
 
 
 
 
20150
  function toolApprovalOn() { try { return localStorage.getItem(TOOL_APPROVAL_KEY) !== 'off'; } catch { return true; } }
20151
  function toolNeedsApproval(name, tool, gateOn, approved) {
20152
  return !!(tool && (tool._custom || tool._mcp || tool.approvalHost)) && !!gateOn && !approved.has(name);
@@ -20159,6 +20190,14 @@ self.addEventListener('message', async (e) => {
20159
  toolApprovalToggle.checked = toolApprovalOn();
20160
  toolApprovalToggle.addEventListener('change', () => { try { localStorage.setItem(TOOL_APPROVAL_KEY, toolApprovalToggle.checked ? 'on' : 'off'); } catch {} });
20161
  }
 
 
 
 
 
 
 
 
20162
  const keylessWebToggle = document.getElementById('keylessWebToggle');
20163
  if (keylessWebToggle) {
20164
  keylessWebToggle.checked = keylessWebOn();
@@ -20237,7 +20276,7 @@ self.addEventListener('message', async (e) => {
20237
  clearInterval(watch);
20238
  card.classList.add('decided');
20239
  card.replaceChildren(Object.assign(document.createElement('div'), {
20240
- textContent: v === 'deny' ? `Declined: ${name} did not run.` : `Allowed: ${name}${v === 'always' ? ' (for the rest of this session)' : ''}.`,
20241
  }));
20242
  resolve(v);
20243
  }
@@ -20245,9 +20284,9 @@ self.addEventListener('message', async (e) => {
20245
  });
20246
  }
20247
  async function executeToolWithApproval(tc, tool, ctx) {
20248
- if (toolNeedsApproval(tc.name, tool, toolApprovalOn(), toolApprovedThisSession)) {
20249
  const v = await askToolApproval(tc.name, tool, tc.arguments);
20250
- if (v === 'always') toolApprovedThisSession.add(tc.name);
20251
  if (v === 'deny') return toolDeclinedResult(tc.name);
20252
  }
20253
  return tool.execute(tc.arguments, ctx);
@@ -21336,11 +21375,14 @@ self.addEventListener('message', async (e) => {
21336
 
21337
  async function voiceStartRecording() {
21338
  if (VM.recording || VM.state === 'thinking' || VM.state === 'speaking') return;
21339
- if (VM.needsDownload) { await voiceLoadModels(); return; } // first tap = consent to download
 
21340
  // Stop any read-back that's still playing (barge-in on a new turn is allowed pre-record).
21341
  ttsStopAll();
21342
- try { await sttEnsureWorker(); }
21343
- catch (e) { voiceSetState('idle', 'Voice recognizer failed to load: ' + (e.message || e)); return; }
 
 
21344
  try { VM.stream = await navigator.mediaDevices.getUserMedia({ audio: true }); }
21345
  catch (e) { voiceSetState('idle', 'Mic access denied — you can still type below.'); return; }
21346
  VM.chunks = [];
@@ -21351,7 +21393,7 @@ self.addEventListener('message', async (e) => {
21351
  VM.recording = true;
21352
  const t = voiceEl('voiceTranscript'); if (t) t.textContent = '';
21353
  voiceSetState('listening', 'Listening… release to send.');
21354
- VM.partialTimer = setInterval(voiceDoPartial, 1200);
21355
  }
21356
 
21357
  async function voiceDoPartial() {
@@ -21379,6 +21421,7 @@ self.addEventListener('message', async (e) => {
21379
  const blob = new Blob(VM.chunks, { type: (VM.recorder && VM.recorder.mimeType) || 'audio/webm' });
21380
  VM.recorder = null;
21381
  if (!blob.size) { voiceSetState('idle', 'No audio captured — hold the mic and speak.'); return; }
 
21382
  voiceSetState('thinking', 'Transcribing…');
21383
  let text = '';
21384
  try { const pcm = await decodeToWhisperPCM(blob); text = (await voiceTranscribe(pcm)).trim(); }
@@ -21396,8 +21439,96 @@ self.addEventListener('message', async (e) => {
21396
  if (VM.state !== 'speaking') voiceSetState('idle', 'Ready — hold the mic to talk.');
21397
  }
21398
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
21399
  async function onEnterVoiceMode() {
21400
  voiceSetState('idle');
 
 
 
 
 
 
21401
  const [sttC, ttsC] = await Promise.all([sttModelCached(), ttsModelCached()]);
21402
  VM.needsDownload = !(sttC && ttsC);
21403
  if (VM.needsDownload) {
@@ -21512,7 +21643,11 @@ self.addEventListener('message', async (e) => {
21512
  // Auto-read the reply: ALWAYS in Voice mode (v1.1), else only if the user
21513
  // enabled auto-play. Skip (and reset the mic) if the language isn't voiced.
21514
  const voiceReadback = (activeMode === 'voice');
21515
- if (response && response.trim() && ttsVoiceForText(response, 'af_heart') && (voiceReadback || ttsSettings().autoplay)) {
 
 
 
 
21516
  const sb = bubble.querySelector('.speak-btn:not(:disabled)');
21517
  if (sb) speakBubble(sb, response);
21518
  } else if (voiceReadback) {
@@ -23967,7 +24102,7 @@ self.addEventListener('message', async (e) => {
23967
  */
23968
  const THINK_CLOSE = -2;
23969
 
23970
- export const ANYFORMER_URL = 'https://cdn.jsdelivr.net/npm/anyformer@0.4.13/index.js';
23971
 
23972
  // anyformer's progress events → the engine events the SSD worker posts. The llama.cpp path reports 'download' (bytes
23973
  // into OPFS) and 'load' / 'plan'; a hand engine (diskformer back end) reports its own engine events, 'weights' with
@@ -23997,6 +24132,7 @@ export class AnyformerModel {
23997
  static file = undefined; // a GGUF in the repo; undefined lets anyformer pick (the file diskformer checked)
23998
  static budget = 4e9; // GPU bytes: the llama.cpp path plans its expert slots from it; for a hand engine, the fit check
23999
  static poolBytes = undefined; // a hand engine's expert cache (LocalMind's own engines ran with 4 GiB)
 
24000
 
24001
  static async load(_id, opts = {}) {
24002
  const { load, MODELS } = await import(ANYFORMER_URL);
@@ -24005,6 +24141,7 @@ export class AnyformerModel {
24005
  const session = await load(entry || this.repo, {
24006
  ...(opts.budget || !entry ? { budget: opts.budget || this.budget } : {}), maxCtx: opts.maxCtx, signal: opts.signal, ...(this.file ? { file: this.file } : {}),
24007
  ...(this.poolBytes ? { poolBytes: this.poolBytes } : {}), ...(opts.localFile ? { localFile: opts.localFile } : {}),
 
24008
  onProgress: (e) => { const ev = toEngineProgress(e, files); if (ev && opts.onProgress) opts.onProgress(ev); },
24009
  });
24010
  return new this(session);
@@ -24025,20 +24162,22 @@ export class AnyformerModel {
24025
 
24026
  // A hand engine's chunks pass through. On the llama.cpp path: the reply text without its <think> / </think> tags; when
24027
  // </think> first appears, the text up to it goes out, then the same text again with token THINK_CLOSE (an empty
24028
- // delta), so the worker closes the thought with no text lost.
24029
- async *generate(messages, { maxNewTokens = 1024, signal, enableThinking = false } = {}) {
 
 
24030
  if (this.think) {
24031
- for await (const c of this.session.generate(messages, { maxNewTokens, signal, thinking: enableThinking })) yield c;
24032
  return;
24033
  }
24034
  let closed = false;
24035
- for await (const c of this.session.generate(messages, { maxNewTokens, signal, thinking: enableThinking })) {
24036
  const raw = String(c.text || '').replace(/^\s*<think>\n?/, '');
24037
  const at = raw.indexOf('</think>');
24038
  // The worker posts no text on its close event, so the thought's last text goes out first.
24039
  if (at >= 0 && !closed) { closed = true; yield { text: raw.slice(0, at) }; yield { text: raw.slice(0, at), token: THINK_CLOSE }; }
24040
  const text = at >= 0 ? raw.slice(0, at) + raw.slice(at + '</think>'.length).replace(/^\n+/, '') : raw;
24041
- yield { text, tokens: c.tokens, tokPerSec: c.tokPerSec };
24042
  }
24043
  }
24044
 
@@ -24125,11 +24264,13 @@ export class Lfm25_24B extends AnyformerModel {
24125
  // LiquidAI's LFM3-Omni-2B (LiquidAI/LFM3-Omni-2B-ONNX: q4f16 decoder, int8 vision encoder, fp32 audio encoder, 2.14 GB)
24126
  // by anyformer's catalog entry, on diskformer's onnxruntime-web engine (lfm3_omni.js, gated 2026-10-30: 32 of 32
24127
  // greedy tokens equal Liquid's CPU reference for text, image and speech in). transformers.js 4.3.2 has no lfm3_omni
24128
- // class; when it adds one, anyformer moves the entry and LocalMind changes nothing. Text out only for now.
 
24129
  // mediaParts: the SSD worker passes image parts as { type:'image', rgb, width, height } and audio parts as
24130
  // { type:'audio', pcm } (16 kHz mono, at most 30 s); this class turns them into anyformer's content blocks.
24131
  export class Lfm3Omni2B extends AnyformerModel {
24132
  static entry = 'LFM3_OMNI_2B';
 
24133
  get mediaParts() { return true; }
24134
  async *generate(messages, opts) {
24135
  const blocks = messages.map((m) => !Array.isArray(m.content) ? m : {
@@ -32371,7 +32512,7 @@ const GGUF_FILE = 'LFM2.5-230M-Q4_0.gguf';
32371
  // The model loads through anyformer (convergence plan, 2026-10-22): anyformer resolves the repo to diskformer's copy of
32372
  // the LFM2.5 engine (Lfm2Mobile) and returns a session with the same generate / reset / dispose. Pinned to an exact
32373
  // version; LocalMind's anyformer_models.js pins the same one.
32374
- const ANYFORMER_URL = 'https://cdn.jsdelivr.net/npm/anyformer@0.4.13/index.js';
32375
 
32376
  let anyformerLoad = null;
32377
  let model = null;
 
866
  height: 64px;
867
  }
868
 
869
+ .voice-omni-toggle { margin-left: 10px; font-weight: 400; font-size: 0.78rem; cursor: pointer; white-space: nowrap; }
870
  .attachment-chip.audio-chip {
871
  width: 180px;
872
  height: 48px;
 
2502
  <li><strong>Artifact preview</strong> &mdash; <code>```html</code> / <code>```svg</code> / <code>```artifact</code> code blocks get a live sandboxed iframe below the code (<code>sandbox=&quot;allow-scripts&quot;</code>, no same-origin). Safe to run model-generated UI inline.</li>
2503
  <li><strong>Voice to text</strong> &mdash; 🗣 button left of the input records mic audio, decodes to 16 kHz mono PCM on-device, and runs Whisper-base via WebGPU to transcribe into the input. ~80 MB first-use download; nothing leaves the device.</li>
2504
  <li><strong>Read aloud</strong> &mdash; 🔊 button under any reply speaks it via Kokoro-82M (runs on WASM, so it never competes with the chat model). Streams sentence-by-sentence; English voices in Settings, and a Hindi reply uses the device&rsquo;s own Hindi voice when it has one. ~80 MB first-use download; nothing leaves the device.</li>
2505
+ <li><strong>Voice mode</strong> &mdash; 🎙 chip opens a hands-free loop: hold the mic (or hold <strong>V</strong>) to talk, release to send. Moonshine transcribes on-device (English; other languages fall back to Whisper), the model answers, and Kokoro reads the reply back. Turn-based; the chat model stays loaded. ~180 MB first-use download. <strong>Omni voice</strong> (the switch in the Voice panel) swaps all three for LFM3-Omni 2B: it hears the recording itself (no transcription, so no misheard words in between) and answers in its own voice; it switches the chat model to LFM3-Omni and back when you turn it off.</li>
2506
  <li><strong>Python code tool</strong> (agent-capable models) &mdash; model can call <code>run_python</code> to execute Python in a sandboxed Pyodide worker. numpy / pandas / matplotlib auto-install on import. ~10 MB first-use download.</li>
2507
  <li><strong>Cache management</strong> &mdash; view/clear cached models in Settings</li>
2508
  <li><strong>Custom models</strong> &mdash; Settings &rarr; paste a Hugging Face ONNX repo id (causal LMs, runs on WebGPU) <em>or a direct GGUF URL</em> (runs via wllama, WebGPU when present / CPU otherwise). ONNX repos are probed via the HF API for the best quant; oversized models warn (but aren&rsquo;t blocked) so you can try on a capable GPU. GGUF URLs are sized via a HEAD request; point at the first shard of a split GGUF (&hellip;-00001-of-&hellip;) for files over wllama&rsquo;s 2 GB per-file / ~4 GB total limits.</li>
 
2681
  <li><strong>Gemma 4 E2B &middot; WebGPU kernels</strong> (~2 GB) &mdash; text + agent. Google's Gemma 4 E2B (QAT mobile, int4) on a from-scratch WebGPU engine: hand-written WGSL for QAT int4 matmul, embed-gather-norm, RoPE/RMSNorm and GQA + sliding-window attention, reading the QAT-mobile weights directly. Ported from <em>webml-community/gemma-4-webgpu-kernels</em>. WebGPU-only. About 4&times; the ONNX Gemma 4 E2B&rsquo;s decode speed on the same Mac (171 vs 42 tok/s, M4 Pro); text-only.</li>
2682
  <li><strong>Gemma 4 E4B &middot; WebGPU kernels</strong> (~3.6 GB) &mdash; text + agent. The larger Gemma 4 (QAT mobile) on the same hand-written WebGPU engine. Its per-layer embedding table can also live on disk (Settings &rarr; Models, below).</li>
2683
  <li><strong>LFM3-VL 2B</strong> (~1.9 GB) &mdash; text and images, with reasoning. Liquid AI's LFM3-VL-2B (October 2026): the LFM3 language model with a new vision encoder, through Transformers.js on WebGPU. Attach an image and ask about it. Show reasoning turns its step-by-step thinking on.</li>
2684
+ <li><strong>LFM3-Omni 2B</strong> (~2.1 GB) &mdash; text, images and speech in, with reasoning and tools. Liquid AI's LFM3-Omni-2B (October 2026), one model that reads text, pictures and spoken questions (up to 30 seconds a clip). Transformers.js does not run it yet, so it comes through anyformer on diskformer's own onnxruntime-web engine; the first load stores the files in this browser once. In chat it answers in text; in Voice mode, turn on <strong>Omni voice</strong> and it hears your recording directly and answers in its own voice (the speech part adds about 0.5 GB to the download).</li>
2685
  <li><strong>Gemma 4 E2B</strong> (~1.5 GB) &mdash; multimodal (image + audio) + agent.</li>
2686
  <li><strong>Gemma 4 E4B</strong> (~4.9 GB) &mdash; multimodal + agent, best quality.</li>
2687
  <li><strong>LFM2.5 230M &middot; GGUF</strong> (~170 MB) &mdash; the same LFM2.5 230M through wllama (llama.cpp &rarr; WebAssembly), reading a Q5_K_M GGUF. WebGPU when your browser has it, CPU otherwise, so it also runs where WebGPU is missing.</li>
 
2851
  <input type="checkbox" id="toolApprovalToggle" checked>
2852
  Ask before running custom and MCP tools <span style="font-weight:400;color:var(--gray-500)">(they send the model's arguments to another server; also asks before weather and Wikipedia lookups)</span>
2853
  </label>
2854
+ <div class="toggle-row" id="toolAlwaysRow" style="display:none">
2855
+ <span>Always allowed: <span id="toolAlwaysList"></span></span>
2856
+ <button type="button" class="btn-icon" id="toolAlwaysForget">Forget</button>
2857
+ </div>
2858
  <label class="toggle-row" id="keylessWebRow">
2859
  <input type="checkbox" id="keylessWebToggle" checked>
2860
  Keyless web tools: weather and Wikipedia <span style="font-weight:400;color:var(--gray-500)">(agent-capable models; sends the place name to open-meteo.com or the topic to wikipedia.org, no key or account)</span>
 
3010
  <option value="lfm3-1.7b">LFM3 1.7B &middot; Agent (~1.4 GB)</option>
3011
  <option value="gemini-nano">Gemini Nano &middot; Built into Chrome (0 MB)</option>
3012
  <option value="lfm3-vl-2b">LFM3-VL 2B &middot; Text and images (~1.9 GB)</option>
3013
+ <option value="lfm3-omni-2b-anyformer">LFM3-Omni 2B &middot; Hears and speaks (~2.6 GB)</option>
3014
  <option value="gemma4-e2b">Gemma 4 E2B &middot; Multimodal (~1.5 GB)</option>
3015
  <option value="gemma4-e4b">Gemma 4 E4B &middot; Multimodal (~4.9 GB)</option>
3016
  <option value="lfm2-230m-gguf">LFM2.5 230M &middot; GGUF (~170 MB)</option>
 
3175
  </div>
3176
  </div>
3177
  <div class="mode-extra" id="voicePanel" role="region" aria-label="Voice conversation">
3178
+ <div class="mode-extra-head">Voice &middot; <span class="memory-count" id="voiceEngineLabel">Moonshine STT + Kokoro TTS &middot; on-device</span>
3179
+ <label class="voice-omni-toggle" title="LFM3-Omni 2B hears your recording directly and answers in its own voice: no transcription step, no separate voice model. Switches the chat model to LFM3-Omni (~2.6 GB once); off switches back.">
3180
+ <input type="checkbox" id="omniVoiceToggle"> Omni voice
3181
+ </label>
3182
+ </div>
3183
  <div class="voice-stage">
3184
  <button type="button" class="voice-mic" id="voiceMicBtn" aria-label="Hold to talk, release to send">
3185
  <span class="voice-mic-icon" id="voiceMicIcon">&#127897;&#65039;</span>
 
8040
  // The model loads through anyformer (convergence plan, 2026-10-22): anyformer resolves the repo to diskformer's copy
8041
  // of the engine (Gemma4Mobile) and, with ple, builds the OPFS per-layer-embedding table itself. The worker then drives
8042
  // session.engine exactly as it drove its own engine. Pinned to the same anyformer as anyformer_models.js.
8043
+ const ANYFORMER_URL = 'https://cdn.jsdelivr.net/npm/anyformer@0.4.14/index.js';
8044
 
8045
  let anyformerLoad = null;
8046
  let session = null;
 
8185
  // rolled-in module URLs (moduleUrl()).
8186
  // The model loads through anyformer (convergence plan, 2026-10-22): anyformer resolves the repo to diskformer's copy of
8187
  // the engine (TernaryBonsai2); the worker drives session.engine as before and attaches DFlash / MTP itself.
8188
+ const ANYFORMER_URL = 'https://cdn.jsdelivr.net/npm/anyformer@0.4.14/index.js';
8189
  const DFLASH_URL = '__TERNARY_BONSAI2_DFLASH_URL__';
8190
  const MTP_URL = '__TERNARY_BONSAI2_MTP_URL__';
8191
  // MTP speculative decoding (Settings → Models, default OFF; posted as `mtp` + `mtpDepth` on the
 
8708
  try { model.reset(); } catch (_) {}
8709
  }
8710
  lastHistoryLen = msgs.length;
8711
+ // Speech out (LFM3-Omni, Voice mode's Omni toggle): the engine sets up Liquid's speech-to-speech prompt (its own
8712
+ // system line and <|mixed_start|>) only when there is no system turn. With LocalMind's system prompt (tool
8713
+ // instructions, ~1,900 tokens) in front, the first in-app run answered a spoken question about mountains with a
8714
+ // refusal about weather tools (2026-10-30), so a spoken turn goes without it: no tools in Omni voice for now.
8715
+ const speech = d.speech === true && model.mediaParts === true;
8716
+ if (speech) while (msgs.length && msgs[0].role === 'system') msgs.shift();
8717
  abortController = new AbortController();
8718
  const open = model.thinkOpenTokenId, close = model.thinkCloseTokenId;
8719
  // A reasoning-only model (LFM2.5 24B: alwaysThinks) opens <think> in every prompt, so its reply always starts as reasoning.
 
8725
  maxNewTokens: cfg.max_new_tokens || cfg.max_tokens || 1024,
8726
  signal: abortController.signal,
8727
  enableThinking: wantThink,
8728
+ ...(speech ? { speech: true } : {}),
8729
  });
8730
+ let toolMarkup = false;
8731
  for await (const out of stream) {
8732
  if (stopFlag) break;
8733
  const full = out && typeof out.text === 'string' ? out.text : '';
8734
  last = full;
8735
+ // The model's speech, until a tool call starts (the interleave writes its text before its sound, so the
8736
+ // markup is never voiced); the page plays it.
8737
+ if (speech && !toolMarkup && (full.includes('<tool_call>') || full.includes('<function='))) toolMarkup = true;
8738
+ // A copy: out.audio may be a view of the engine's own buffer, which transferring would detach (2026-10-30).
8739
+ if (speech && !toolMarkup && out.audio && out.audio.length) { const pcm = Float32Array.from(out.audio); self.postMessage({ type: 'speech', pcm, sampleRate: out.sampleRate || 24000 }, [pcm.buffer]); }
8740
  // A multi-byte character split across tokens decodes as a trailing U+FFFD until its last byte
8741
  // arrives. Hold the delta back until then: posting it sent '\uFFFD' and then cut the next delta
8742
  // one code unit late, so 🍂 arrived as '\uFFFD\uDF42' (2026-10-14).
 
9510
  id: 'LiquidAI/LFM3-Omni-2B-ONNX',
9511
  label: 'LFM3-Omni 2B',
9512
  backend: 'anyformer-lfm3-omni',
9513
+ size: '~2.6 GB',
9514
  type: 'causal',
9515
  multimodal: true,
9516
  imageInput: true,
 
19057
  }
19058
  }
19059
 
19060
+ else if (data.type === 'speech') {
19061
+ if (generating && activeMode === 'voice') omniSpeechChunk(data.pcm, data.sampleRate);
19062
+ }
19063
+
19064
  else if (data.type === 'token') {
19065
  // Always accumulate so non-streaming callers (chat UI, non-streaming
19066
  // API) still receive the full text via the 'complete' handler.
 
19873
  attachments: attData || null,
19874
  enableThinking,
19875
  generationConfig: genConfig,
19876
+ speech: typeof omniVoiceActive === 'function' && omniVoiceActive(),
19877
  });
19878
  });
19879
  }
 
20170
  // a flag for whether any tool was actually called.
20171
  // TOOL-APPROVAL (Open Q#11): custom and MCP tools send the model's arguments to another server, so each call
20172
  // waits for Allow / Always / Deny on a card in the chat, unless the gate is off (Settings → Tools) or the user
20173
+ // chose "Always" for that tool. "Always" is stored (lm_tool_always, a JSON list of tool names), so it holds across
20174
+ // chats and reloads until Settings → Tools → Forget. The keyless web built-ins (get_weather, wikipedia_lookup)
20175
+ // carry an approvalHost and ask the same way; every other built-in runs without a card.
20176
+ const TOOL_APPROVAL_KEY = 'lm_tool_approval', TOOL_ALWAYS_KEY = 'lm_tool_always';
20177
+ function toolAlwaysLoad() { try { const a = JSON.parse(localStorage.getItem(TOOL_ALWAYS_KEY) || '[]'); return new Set(Array.isArray(a) ? a.filter((x) => typeof x === 'string') : []); } catch { return new Set(); } }
20178
+ const toolAlwaysAllowed = toolAlwaysLoad();
20179
+ function toolAlwaysAdd(name) { toolAlwaysAllowed.add(name); try { localStorage.setItem(TOOL_ALWAYS_KEY, JSON.stringify([...toolAlwaysAllowed])); } catch {} }
20180
+ function toolAlwaysClear() { toolAlwaysAllowed.clear(); try { localStorage.removeItem(TOOL_ALWAYS_KEY); } catch {} }
20181
  function toolApprovalOn() { try { return localStorage.getItem(TOOL_APPROVAL_KEY) !== 'off'; } catch { return true; } }
20182
  function toolNeedsApproval(name, tool, gateOn, approved) {
20183
  return !!(tool && (tool._custom || tool._mcp || tool.approvalHost)) && !!gateOn && !approved.has(name);
 
20190
  toolApprovalToggle.checked = toolApprovalOn();
20191
  toolApprovalToggle.addEventListener('change', () => { try { localStorage.setItem(TOOL_APPROVAL_KEY, toolApprovalToggle.checked ? 'on' : 'off'); } catch {} });
20192
  }
20193
+ function renderToolAlways() {
20194
+ const row = document.getElementById('toolAlwaysRow'), list = document.getElementById('toolAlwaysList');
20195
+ if (!row || !list) return;
20196
+ list.textContent = [...toolAlwaysAllowed].join(', ');
20197
+ row.style.display = toolAlwaysAllowed.size ? '' : 'none';
20198
+ }
20199
+ renderToolAlways();
20200
+ document.getElementById('toolAlwaysForget')?.addEventListener('click', () => { toolAlwaysClear(); renderToolAlways(); });
20201
  const keylessWebToggle = document.getElementById('keylessWebToggle');
20202
  if (keylessWebToggle) {
20203
  keylessWebToggle.checked = keylessWebOn();
 
20276
  clearInterval(watch);
20277
  card.classList.add('decided');
20278
  card.replaceChildren(Object.assign(document.createElement('div'), {
20279
+ textContent: v === 'deny' ? `Declined: ${name} did not run.` : `Allowed: ${name}${v === 'always' ? ' (always; Settings → Tools forgets it)' : ''}.`,
20280
  }));
20281
  resolve(v);
20282
  }
 
20284
  });
20285
  }
20286
  async function executeToolWithApproval(tc, tool, ctx) {
20287
+ if (toolNeedsApproval(tc.name, tool, toolApprovalOn(), toolAlwaysAllowed)) {
20288
  const v = await askToolApproval(tc.name, tool, tc.arguments);
20289
+ if (v === 'always') { toolAlwaysAdd(tc.name); renderToolAlways(); }
20290
  if (v === 'deny') return toolDeclinedResult(tc.name);
20291
  }
20292
  return tool.execute(tc.arguments, ctx);
 
21375
 
21376
  async function voiceStartRecording() {
21377
  if (VM.recording || VM.state === 'thinking' || VM.state === 'speaking') return;
21378
+ const omni = omniVoiceActive(); // LFM3-Omni hears the clip itself: no Moonshine, no Kokoro
21379
+ if (VM.needsDownload && !omni) { await voiceLoadModels(); return; } // first tap = consent to download
21380
  // Stop any read-back that's still playing (barge-in on a new turn is allowed pre-record).
21381
  ttsStopAll();
21382
+ if (!omni) {
21383
+ try { await sttEnsureWorker(); }
21384
+ catch (e) { voiceSetState('idle', 'Voice recognizer failed to load: ' + (e.message || e)); return; }
21385
+ }
21386
  try { VM.stream = await navigator.mediaDevices.getUserMedia({ audio: true }); }
21387
  catch (e) { voiceSetState('idle', 'Mic access denied — you can still type below.'); return; }
21388
  VM.chunks = [];
 
21393
  VM.recording = true;
21394
  const t = voiceEl('voiceTranscript'); if (t) t.textContent = '';
21395
  voiceSetState('listening', 'Listening… release to send.');
21396
+ if (!omni) VM.partialTimer = setInterval(voiceDoPartial, 1200);
21397
  }
21398
 
21399
  async function voiceDoPartial() {
 
21421
  const blob = new Blob(VM.chunks, { type: (VM.recorder && VM.recorder.mimeType) || 'audio/webm' });
21422
  VM.recorder = null;
21423
  if (!blob.size) { voiceSetState('idle', 'No audio captured — hold the mic and speak.'); return; }
21424
+ if (omniVoiceActive()) { await omniVoiceSend(blob); return; }
21425
  voiceSetState('thinking', 'Transcribing…');
21426
  let text = '';
21427
  try { const pcm = await decodeToWhisperPCM(blob); text = (await voiceTranscribe(pcm)).trim(); }
 
21439
  if (VM.state !== 'speaking') voiceSetState('idle', 'Ready — hold the mic to talk.');
21440
  }
21441
 
21442
+ // ── Omni voice (2026-10-30): LFM3-Omni 2B hears the recording and answers in its own voice ──
21443
+ // The clip goes to the model as an audio attachment (16 kHz, at most 30 s); the worker streams the model's 24 kHz
21444
+ // speech as 'speech' messages, played through Kokoro's audio context and registered in TTS.sessions, so the mic
21445
+ // (ttsStopAll) and leaving Voice mode stop it like a read-back. Kokoro does not read a reply the model spoke.
21446
+ const OMNI_VOICE_KEY = 'lm_omni_voice', OMNI_MODEL_KEY = 'lfm3-omni-2b-anyformer', OMNI_MAX_SECONDS = 30;
21447
+ let omniPrevModel = null;
21448
+ function omniVoiceOn() { try { return localStorage.getItem(OMNI_VOICE_KEY) === 'on'; } catch { return false; } }
21449
+ function omniVoiceActive() { return activeMode === 'voice' && omniVoiceOn() && activeModelKey === OMNI_MODEL_KEY && modelReady; }
21450
+ function omniVoiceLabel() {
21451
+ const l = voiceEl('voiceEngineLabel');
21452
+ if (l) l.innerHTML = omniVoiceOn() ? 'LFM3-Omni 2B hears and speaks &middot; on-device' : 'Moonshine STT + Kokoro TTS &middot; on-device';
21453
+ const t = voiceEl('omniVoiceToggle'); if (t) t.checked = omniVoiceOn();
21454
+ }
21455
+ function omniPickModel(key) {
21456
+ if (!key || (key === activeModelKey && modelReady)) return;
21457
+ if (generating) { setTimeout(() => omniPickModel(key), 500); return; } // the switch waits for the reply in flight
21458
+ modelSelect.value = key;
21459
+ modelSelect.dispatchEvent(new Event('change', { bubbles: true }));
21460
+ }
21461
+ function omniVoiceSet(on) {
21462
+ try { localStorage.setItem(OMNI_VOICE_KEY, on ? 'on' : 'off'); } catch {}
21463
+ omniVoiceLabel();
21464
+ if (on && activeModelKey !== OMNI_MODEL_KEY) { omniPrevModel = activeModelKey; omniPickModel(OMNI_MODEL_KEY); }
21465
+ else if (!on && activeModelKey === OMNI_MODEL_KEY && omniPrevModel) { const k = omniPrevModel; omniPrevModel = null; omniPickModel(k); }
21466
+ if (activeMode === 'voice') voiceSetState('idle', on ? 'Hold the mic (or press and hold V) to talk. LFM3-Omni hears you and answers in its own voice.' : 'Ready — hold the mic to talk.');
21467
+ }
21468
+ voiceEl('omniVoiceToggle')?.addEventListener('change', (e) => omniVoiceSet(e.target.checked));
21469
+ omniVoiceLabel();
21470
+
21471
+ async function omniVoiceSend(blob) {
21472
+ let pcm;
21473
+ try { pcm = await decodeToWhisperPCM(blob); }
21474
+ catch (e) { voiceSetState('idle', 'Could not read the recording: ' + (e.message || e)); return; }
21475
+ const secs = pcm.length / 16000;
21476
+ if (secs > OMNI_MAX_SECONDS) { voiceSetState('idle', `That clip is ${secs.toFixed(0)} s; LFM3-Omni takes up to ${OMNI_MAX_SECONDS} s. Try a shorter question.`); return; }
21477
+ const t = voiceEl('voiceTranscript'); if (t) t.textContent = `🎙 ${secs.toFixed(1)} s clip`;
21478
+ addAttachment({ type: 'audio', blob, pcmData: pcm, thumb: URL.createObjectURL(blob), name: 'voice clip' });
21479
+ chatInput.value = '';
21480
+ voiceSetState('thinking', 'Listening to your clip…');
21481
+ try { await sendMessage(); }
21482
+ catch (e) { voiceSetState('idle', 'Send failed: ' + (e.message || e)); return; }
21483
+ if (VM.state !== 'speaking') voiceSetState('idle', 'Ready — hold the mic to talk.');
21484
+ }
21485
+
21486
+ // One playback session per reply: chunks scheduled back to back on Kokoro's audio context.
21487
+ const OMNI_SPEECH = { session: null };
21488
+ function omniSpeechChunk(pcm, sr) {
21489
+ let sp = OMNI_SPEECH.session;
21490
+ if (!sp) {
21491
+ const ctx = ttsGetCtx();
21492
+ try { if (ctx.state === 'suspended') { ctx.resume().catch(() => {}); ttsArmResumeOnGesture(ctx); } } catch (_) {}
21493
+ const id = ++TTS.seq, sources = [];
21494
+ let nextTime = 0, pending = 0, ended = false, done = false;
21495
+ const finish = () => {
21496
+ if (done) return; done = true;
21497
+ TTS.sessions.delete(id); if (TTS.active && TTS.active.id === id) TTS.active = null;
21498
+ if (OMNI_SPEECH.session === sp) OMNI_SPEECH.session = null;
21499
+ if (activeMode === 'voice' && !generating) voiceSetState('idle', 'Ready — hold the mic to talk.');
21500
+ };
21501
+ sp = {
21502
+ sawAudio: false,
21503
+ onChunk(p, rate) {
21504
+ if (done) return;
21505
+ const buf = ctx.createBuffer(1, p.length, rate);
21506
+ buf.copyToChannel(p, 0);
21507
+ const node = ctx.createBufferSource();
21508
+ node.buffer = buf; node.connect(ctx.destination);
21509
+ const at = Math.max(ctx.currentTime + 0.02, nextTime);
21510
+ node.start(at); nextTime = at + buf.duration;
21511
+ pending++; sources.push(node); sp.sawAudio = true;
21512
+ node.onended = () => { pending--; if (ended && pending <= 0) finish(); };
21513
+ if (activeMode === 'voice') voiceSetState('speaking', 'Speaking…');
21514
+ },
21515
+ end() { ended = true; if (pending <= 0) finish(); },
21516
+ stop() { sources.forEach((n) => { try { n.stop(); } catch (_) {} }); finish(); },
21517
+ };
21518
+ OMNI_SPEECH.session = sp;
21519
+ TTS.sessions.set(id, sp); TTS.active = { id, session: sp, button: null };
21520
+ }
21521
+ sp.onChunk(pcm instanceof Float32Array ? pcm : new Float32Array(pcm), sr || 24000);
21522
+ }
21523
+
21524
  async function onEnterVoiceMode() {
21525
  voiceSetState('idle');
21526
+ omniVoiceLabel();
21527
+ if (omniVoiceOn()) {
21528
+ if (activeModelKey !== OMNI_MODEL_KEY) { omniPrevModel = activeModelKey; omniPickModel(OMNI_MODEL_KEY); }
21529
+ voiceSetState('idle', 'Hold the mic (or press and hold V) to talk. LFM3-Omni hears you and answers in its own voice.');
21530
+ return;
21531
+ }
21532
  const [sttC, ttsC] = await Promise.all([sttModelCached(), ttsModelCached()]);
21533
  VM.needsDownload = !(sttC && ttsC);
21534
  if (VM.needsDownload) {
 
21643
  // Auto-read the reply: ALWAYS in Voice mode (v1.1), else only if the user
21644
  // enabled auto-play. Skip (and reset the mic) if the language isn't voiced.
21645
  const voiceReadback = (activeMode === 'voice');
21646
+ const omniSpoke = !!(OMNI_SPEECH.session && OMNI_SPEECH.session.sawAudio);
21647
+ if (OMNI_SPEECH.session) OMNI_SPEECH.session.end();
21648
+ if (omniSpoke) {
21649
+ // LFM3-Omni already said it; the 🔊 button still reads the text with Kokoro on request.
21650
+ } else if (response && response.trim() && ttsVoiceForText(response, 'af_heart') && (voiceReadback || ttsSettings().autoplay)) {
21651
  const sb = bubble.querySelector('.speak-btn:not(:disabled)');
21652
  if (sb) speakBubble(sb, response);
21653
  } else if (voiceReadback) {
 
24102
  */
24103
  const THINK_CLOSE = -2;
24104
 
24105
+ export const ANYFORMER_URL = 'https://cdn.jsdelivr.net/npm/anyformer@0.4.14/index.js';
24106
 
24107
  // anyformer's progress events → the engine events the SSD worker posts. The llama.cpp path reports 'download' (bytes
24108
  // into OPFS) and 'load' / 'plan'; a hand engine (diskformer back end) reports its own engine events, 'weights' with
 
24132
  static file = undefined; // a GGUF in the repo; undefined lets anyformer pick (the file diskformer checked)
24133
  static budget = 4e9; // GPU bytes: the llama.cpp path plans its expert slots from it; for a hand engine, the fit check
24134
  static poolBytes = undefined; // a hand engine's expert cache (LocalMind's own engines ran with 4 GiB)
24135
+ static speech = false; // load the speech graphs too (LFM3-Omni: depth and codec decoders, 515 MB)
24136
 
24137
  static async load(_id, opts = {}) {
24138
  const { load, MODELS } = await import(ANYFORMER_URL);
 
24141
  const session = await load(entry || this.repo, {
24142
  ...(opts.budget || !entry ? { budget: opts.budget || this.budget } : {}), maxCtx: opts.maxCtx, signal: opts.signal, ...(this.file ? { file: this.file } : {}),
24143
  ...(this.poolBytes ? { poolBytes: this.poolBytes } : {}), ...(opts.localFile ? { localFile: opts.localFile } : {}),
24144
+ ...(this.speech ? { speech: true } : {}),
24145
  onProgress: (e) => { const ev = toEngineProgress(e, files); if (ev && opts.onProgress) opts.onProgress(ev); },
24146
  });
24147
  return new this(session);
 
24162
 
24163
  // A hand engine's chunks pass through. On the llama.cpp path: the reply text without its <think> / </think> tags; when
24164
  // </think> first appears, the text up to it goes out, then the same text again with token THINK_CLOSE (an empty
24165
+ // delta), so the worker closes the thought with no text lost. With speech (LFM3-Omni) a chunk also carries `audio`,
24166
+ // the new 24 kHz PCM since the last chunk, and `sampleRate`.
24167
+ async *generate(messages, { maxNewTokens = 1024, signal, enableThinking = false, speech = false } = {}) {
24168
+ const opts = { maxNewTokens, signal, thinking: enableThinking, ...(speech ? { speech: true } : {}) };
24169
  if (this.think) {
24170
+ for await (const c of this.session.generate(messages, opts)) yield c;
24171
  return;
24172
  }
24173
  let closed = false;
24174
+ for await (const c of this.session.generate(messages, opts)) {
24175
  const raw = String(c.text || '').replace(/^\s*<think>\n?/, '');
24176
  const at = raw.indexOf('</think>');
24177
  // The worker posts no text on its close event, so the thought's last text goes out first.
24178
  if (at >= 0 && !closed) { closed = true; yield { text: raw.slice(0, at) }; yield { text: raw.slice(0, at), token: THINK_CLOSE }; }
24179
  const text = at >= 0 ? raw.slice(0, at) + raw.slice(at + '</think>'.length).replace(/^\n+/, '') : raw;
24180
+ yield { text, tokens: c.tokens, tokPerSec: c.tokPerSec, ...(c.audio ? { audio: c.audio, sampleRate: c.sampleRate } : {}) };
24181
  }
24182
  }
24183
 
 
24264
  // LiquidAI's LFM3-Omni-2B (LiquidAI/LFM3-Omni-2B-ONNX: q4f16 decoder, int8 vision encoder, fp32 audio encoder, 2.14 GB)
24265
  // by anyformer's catalog entry, on diskformer's onnxruntime-web engine (lfm3_omni.js, gated 2026-10-30: 32 of 32
24266
  // greedy tokens equal Liquid's CPU reference for text, image and speech in). transformers.js 4.3.2 has no lfm3_omni
24267
+ // class; when it adds one, anyformer moves the entry and LocalMind changes nothing. Speech out from anyformer 0.4.14
24268
+ // (diskformer 0.6.12), used by Voice mode's Omni toggle; text chat stays text out.
24269
  // mediaParts: the SSD worker passes image parts as { type:'image', rgb, width, height } and audio parts as
24270
  // { type:'audio', pcm } (16 kHz mono, at most 30 s); this class turns them into anyformer's content blocks.
24271
  export class Lfm3Omni2B extends AnyformerModel {
24272
  static entry = 'LFM3_OMNI_2B';
24273
+ static speech = true; // Voice mode's Omni toggle speaks with it; the speech graphs load with the model (2.6 GB in all)
24274
  get mediaParts() { return true; }
24275
  async *generate(messages, opts) {
24276
  const blocks = messages.map((m) => !Array.isArray(m.content) ? m : {
 
32512
  // The model loads through anyformer (convergence plan, 2026-10-22): anyformer resolves the repo to diskformer's copy of
32513
  // the LFM2.5 engine (Lfm2Mobile) and returns a session with the same generate / reset / dispose. Pinned to an exact
32514
  // version; LocalMind's anyformer_models.js pins the same one.
32515
+ const ANYFORMER_URL = 'https://cdn.jsdelivr.net/npm/anyformer@0.4.14/index.js';
32516
 
32517
  let anyformerLoad = null;
32518
  let model = null;