NakliTechie commited on
Commit
c0126f8
·
1 Parent(s): d9e4ace

Sync from GitHub 2026-10-28T14:14:37Z

Browse files
Files changed (1) hide show
  1. index.html +65 -28
index.html CHANGED
@@ -2674,7 +2674,7 @@
2674
  <li><strong>Ternary Bonsai 2 27B</strong> (~5.9 GB) &mdash; text + agent. PrismML's Ternary Bonsai 2 27B (Sept 2026; ternary 1.75-bit weights, Qwen3.8-27B backbone, 98% of the full-precision model's benchmark score) on a from-scratch WebGPU engine: hand-written WGSL for ternary dequant + Gated-DeltaNet linear attention, reading the PTQ1_0 GGUF directly. Ported from <em>webml-community/ternary-bonsai-2-webgpu-kernels</em>; Apache-2.0 weights. WebGPU-only &mdash; the biggest in-browser model here. A reasoning model: flip <strong>Show reasoning</strong> in Settings for its chain-of-thought. Text-only here (its optional vision tower is not loaded).</li>
2675
  <li><strong>Underdog Saluki 27B</strong> (~7.9 GB) &mdash; text + agent. ConwayResearch's Underdog Saluki 27B 1.0 (Oct 2026): Qwen3.8-27B as one 2-bit-class IQ2-mix GGUF, tuned to keep tool calling intact (the vendor reports 88 of 120 on a frozen BFCL v4 tool-calling set, against 84 for the full-size model and 70 for Ternary Bonsai 2). Runs on a from-scratch WebGPU engine that reads the GGUF&rsquo;s eleven llama.cpp block types directly; greedy output matches llama.cpp on the same file. The first load copies the 7.9 GB file into this browser&rsquo;s on-device storage once (or pick the file you already have with <strong>Load .gguf from disk&hellip;</strong>); it needs about 8 GB of GPU memory. Apache-2.0 weights. A reasoning model: flip <strong>Show reasoning</strong> in Settings for its chain-of-thought. Text-only here (its optional vision add-on is not loaded).</li>
2676
  <li><strong>Qwen3.5 4B</strong> (~3 GB) &mdash; text + image + agent (tool calling). Qwen3.5 at q4f16, Apache-2.0; its native 262K context is capped to 32K to keep browser memory bounded.</li>
2677
- <li><strong>LFM2.5 1.2B</strong> (~760 MB) &mdash; text + agent (tool calling). Liquid AI's LFM2.5-1.2B-Instruct (Jan 2026) &mdash; same LFM2 hybrid arch as the 230M default, RL-tuned for instruction following + tool use. q4f16, 32K context.</li>
2678
  <li><strong>Gemini Nano</strong> (0 MB) &mdash; Google's on-device model built into Chrome, via the Prompt API. Zero download, no WebGPU. Chrome/Edge only; runs where your browser exposes <code>self.LanguageModel</code>. When Nano is already downloaded on desktop Chrome/Edge, LocalMind starts on it by default so you can chat with no model download.</li>
2679
  <li><strong>LFM2.5 230M &middot; WebGPU kernels</strong> (~140 MB) &mdash; the same LFM2.5 230M run on a from-scratch WebGPU engine whose every kernel (RoPE, RMSNorm, Q4_0 dequant, the LFM2 short-conv, GQA attention) is hand-written WGSL, reading a Q4_0 GGUF directly (no ONNX, no llama.cpp). Ported from <em>webml-community/lfm2-webgpu-kernels</em>. WebGPU-only, tuned for maximal decode throughput.</li>
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>
@@ -2996,7 +2996,7 @@
2996
  <option value="ternary-bonsai-2-27b-webgpu">Ternary Bonsai 2 27B &middot; Agent (~5.9 GB)</option>
2997
  <option value="underdog-saluki-27b">Underdog Saluki 27B &middot; Agent (~7.9 GB)</option>
2998
  <option value="qwen35-4b">Qwen3.5 4B &middot; Vision + Agent (~3 GB)</option>
2999
- <option value="lfm2-1.2b">LFM2.5 1.2B &middot; Agent (~760 MB)</option>
3000
  <option value="gemini-nano">Gemini Nano &middot; Built into Chrome (0 MB)</option>
3001
  <option value="gemma4-e2b">Gemma 4 E2B &middot; Multimodal (~1.5 GB)</option>
3002
  <option value="gemma4-e4b">Gemma 4 E4B &middot; Multimodal (~4.9 GB)</option>
@@ -7483,7 +7483,7 @@ reply({ type: 'worker-ready' });
7483
  // recipe (fused fp16 for WebGPU) are in reference/MDLM-algorithm.md.
7484
 
7485
  import * as ortDefault from 'https://cdn.jsdelivr.net/npm/onnxruntime-web@1.26.0/dist/ort.webgpu.min.mjs';
7486
- import { AutoTokenizer } from 'https://cdn.jsdelivr.net/npm/@huggingface/transformers@4.3.1/+esm';
7487
 
7488
  // Tiny-A2D Qwen3-0.6B MDLM token ids (tokenizer.mask_token_id / <|im_end|>).
7489
  export const MASK_ID = 151669n;
@@ -8810,7 +8810,7 @@ self.onmessage = async (e) => {
8810
  // WASM by default so it never contends with the chat model's WebGPU. On-device
8811
  // only — the sole network touch is the one-time model fetch. (Non-English turns
8812
  // route to the existing Whisper worker on the main thread instead.)
8813
- import { pipeline, env } from 'https://cdn.jsdelivr.net/npm/@huggingface/transformers@4.3.1/+esm';
8814
  env.allowLocalModels = false;
8815
  env.allowRemoteModels = true;
8816
 
@@ -8853,7 +8853,7 @@ self.onmessage = async (e) => {
8853
  // every module fp16 (~1.0 GB); Light = q4f16 (~0.36 GB) for phones. (q4f16 was
8854
  // degenerate on the 256M under 4.2.0; the 500M on this 3.7.1 pin is under test.)
8855
  //
8856
- // VERSION PIN — this vision worker pins transformers.js to 3.7.1, NOT the @4.3.1
8857
  // the chat/OCR workers use. Verified live 2026-08-21: on 4.2.0 SmolVLM's Idefics3
8858
  // processor preprocesses fine but generation is DEGENERATE (pure gibberish); on
8859
  // 3.7.1 the identical model+image+dtype produces a coherent, image-grounded
@@ -8940,7 +8940,7 @@ import {
8940
  ChatterboxModel,
8941
  AutoProcessor,
8942
  Tensor,
8943
- } from 'https://cdn.jsdelivr.net/npm/@huggingface/transformers@4.3.1/+esm';
8944
 
8945
  env.allowLocalModels = true;
8946
  env.localModelPath = '/models/';
@@ -9055,6 +9055,7 @@ self.addEventListener('message', async (e) => {
9055
  'onnx-community/Apertus-v1.1-4B-Instruct-ONNX', 'onnx-community/Qwen3-4B-ONNX',
9056
  'onnx-community/granite-4.0-1b-ONNX-web', 'onnx-community/granite-4.0-micro-ONNX-web',
9057
  'onnx-community/Ternary-Bonsai-4B-ONNX', // first-generation ternary 4B, replaced by Ternary Bonsai 2 4B (2026-10-16)
 
9058
  ];
9059
 
9060
  const MODELS = {
@@ -9077,23 +9078,23 @@ self.addEventListener('message', async (e) => {
9077
  // top_k 20. Apache-2.0, native tool calling.
9078
  genConfig: { temperature: 0.7, top_k: 20, top_p: 0.8, max_new_tokens: 2048, repetition_penalty: 1.05 },
9079
  },
9080
- 'lfm2-1.2b': {
9081
- id: 'LiquidAI/LFM2.5-1.2B-Instruct-ONNX',
9082
- label: 'LFM2.5 1.2B',
9083
  dtype: 'q4f16',
9084
- size: '~760 MB',
9085
  type: 'causal',
9086
  multimodal: false,
9087
  agentCapable: true,
9088
  contextSize: 32768,
9089
- // Liquid AI LFM2.5-1.2B-Instruct (Jan 2026) — the same LFM2 hybrid arch as
9090
- // the 230M default (so the ChatML-template fallback in generateCausal
9091
- // covers it), RL-tuned for instruction following + tool use (IFBench
9092
- // 47.33). A bigger-brain drop-in sibling of the default. First-party ONNX
9093
- // (q4f16). LFM Open License v1.0. Native 128K context capped to 32K.
9094
- // Liquid recommends temperature 0.3 / min_p 0.15 / rep 1.05; min_p isn't
9095
- // exposed by Transformers.js generate(), so top_p approximates.
9096
- genConfig: { temperature: 0.3, top_k: 20, top_p: 0.95, max_new_tokens: 2048, repetition_penalty: 1.05 },
9097
  },
9098
  'gemini-nano': {
9099
  id: 'chrome-builtin/gemini-nano',
@@ -10485,6 +10486,7 @@ self.addEventListener('message', async (e) => {
10485
  const TOKEN_WEIGHTS = {
10486
  qwen: [0.196, 0.549, 0.473, 0.515, 0.933, 0.717, 1.011, 0.488, 0.644, 0.612, 0.615, 0.291, 0.236, 0.371, 2.845],
10487
  lfm2: [0.202, 0.578, 1.495, 2.353, 2.97, 2.964, 2.971, 2.494, 2.966, 2.956, 2.98, 0.485, 0.357, 0.446, 5.534],
 
10488
  gemma: [0.199, 0.618, 0.28, 0.237, 0.659, 0.424, 0.908, 0.25, 0.361, 0.34, 0.331, 0.352, 0.233, 0.412, 2.692],
10489
  minicpm5: [0.212, 0.427, 0.902, 1.846, 3.18, 3.191, 3.124, 1.986, 2.127, 1.452, 2.386, 1.054, 0.278, 0.626, 0.01],
10490
  unknown: [0.244, 0.711, 1.719, 2.706, 3.657, 3.67, 3.593, 2.868, 3.411, 3.4, 3.427, 1.212, 0.411, 0.729, 6.364],
@@ -10502,7 +10504,7 @@ self.addEventListener('message', async (e) => {
10502
  'gemma4-26b-a4b-ssd': 'gemma',
10503
  'ternary-bonsai-2-2b-webgpu': 'qwen',
10504
  'ternary-bonsai-2-4b-webgpu': 'qwen',
10505
- 'lfm2-1.2b': 'lfm2',
10506
  'lfm2-230m-gguf': 'lfm2',
10507
  'lfm2-230m-webgpu': 'lfm2',
10508
  'gemma4-e2b': 'gemma',
@@ -17620,7 +17622,7 @@ import {
17620
  pipeline,
17621
  AutoTokenizer,
17622
  AutoModelForSequenceClassification,
17623
- } from 'https://cdn.jsdelivr.net/npm/@huggingface/transformers@4.3.1/+esm';
17624
 
17625
  env.allowLocalModels = true;
17626
  env.localModelPath = '/models/';
@@ -17931,7 +17933,7 @@ import {
17931
  Sam3TrackerModel,
17932
  AutoProcessor,
17933
  RawImage,
17934
- } from 'https://cdn.jsdelivr.net/npm/@huggingface/transformers@4.3.1/+esm';
17935
 
17936
  env.allowLocalModels = true;
17937
  env.localModelPath = '/models/';
@@ -18072,7 +18074,7 @@ import {
18072
  RawImage,
18073
  TextStreamer,
18074
  InterruptableStoppingCriteria,
18075
- } from 'https://cdn.jsdelivr.net/npm/@huggingface/transformers@4.3.1/+esm';
18076
 
18077
  env.allowLocalModels = true;
18078
  env.localModelPath = '/models/';
@@ -18245,7 +18247,7 @@ import {
18245
  AutoTokenizer,
18246
  WhisperForConditionalGeneration,
18247
  full,
18248
- } from 'https://cdn.jsdelivr.net/npm/@huggingface/transformers@4.3.1/+esm';
18249
 
18250
  env.allowLocalModels = true;
18251
  env.localModelPath = '/models/';
@@ -18724,10 +18726,10 @@ __rdbCleanup();
18724
  // would race the 'load' postMessage from the main thread — it can arrive
18725
  // before the module-body handler is registered and get dropped. Kick off
18726
  // the import eagerly and have the message handler await it.
18727
- let env, AutoTokenizer, AutoModelForCausalLM, AutoProcessor, Gemma4ForConditionalGeneration, Qwen3_5ForConditionalGeneration, load_image, TextStreamer, InterruptableStoppingCriteria;
18728
  let stopping_criteria;
18729
 
18730
- const __modReady = import('https://cdn.jsdelivr.net/npm/@huggingface/transformers@4.3.1/+esm').then((m) => {
18731
  env = m.env;
18732
  AutoTokenizer = m.AutoTokenizer;
18733
  AutoModelForCausalLM = m.AutoModelForCausalLM;
@@ -18737,6 +18739,7 @@ const __modReady = import('https://cdn.jsdelivr.net/npm/@huggingface/transformer
18737
  load_image = m.load_image;
18738
  TextStreamer = m.TextStreamer;
18739
  InterruptableStoppingCriteria = m.InterruptableStoppingCriteria;
 
18740
  env.allowLocalModels = true;
18741
  env.localModelPath = '/models/';
18742
  env.allowRemoteModels = true;
@@ -18756,6 +18759,20 @@ function progressCallback(p) {
18756
  self.postMessage({ type: 'progress', data: p });
18757
  }
18758
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
18759
  async function loadCausal(modelId, dtype) {
18760
  processor = null; tokenizer = null; model = null;
18761
 
@@ -18768,6 +18785,7 @@ async function loadCausal(modelId, dtype) {
18768
  device: 'webgpu',
18769
  progress_callback: progressCallback,
18770
  });
 
18771
 
18772
  // Warmup: compile WebGPU shaders with a single token
18773
  self.postMessage({ type: 'warmup' });
@@ -18790,7 +18808,7 @@ async function loadMultimodal(modelId, dtype, modelClass) {
18790
  ? Qwen3_5ForConditionalGeneration
18791
  : Gemma4ForConditionalGeneration;
18792
  if (!ConditionalGenerationModel) {
18793
- throw new Error('Transformers.js 4.3.1 does not export ' + modelClass);
18794
  }
18795
  model = await ConditionalGenerationModel.from_pretrained(modelId, {
18796
  dtype: dtype,
@@ -18807,8 +18825,10 @@ async function loadMultimodal(modelId, dtype, modelClass) {
18807
  self.postMessage({ type: 'ready' });
18808
  }
18809
 
18810
- async function generateCausal(chatMessages, id, genConfig) {
18811
  stopping_criteria.reset();
 
 
18812
 
18813
  try {
18814
  let inputs;
@@ -18816,6 +18836,7 @@ async function generateCausal(chatMessages, id, genConfig) {
18816
  inputs = tokenizer.apply_chat_template(chatMessages, {
18817
  add_generation_prompt: true,
18818
  return_dict: true,
 
18819
  });
18820
  } catch (tplErr) {
18821
  // The pinned transformers.js jinja can't parse some real templates — e.g.
@@ -18838,6 +18859,7 @@ async function generateCausal(chatMessages, id, genConfig) {
18838
  inputs = tokenizer.apply_chat_template(chatMessages, {
18839
  add_generation_prompt: true,
18840
  return_dict: true,
 
18841
  chat_template: raw,
18842
  });
18843
  salvaged = true;
@@ -18860,10 +18882,25 @@ async function generateCausal(chatMessages, id, genConfig) {
18860
  }
18861
  }
18862
 
 
 
 
 
 
 
 
18863
  const streamer = new TextStreamer(tokenizer, {
18864
  skip_prompt: true,
18865
  skip_special_tokens: true,
18866
  callback_function: (text) => {
 
 
 
 
 
 
 
 
18867
  self.postMessage({ type: 'token', token: text, id });
18868
  },
18869
  });
@@ -18992,7 +19029,7 @@ self.addEventListener('message', async (e) => {
18992
  if (loadedType === 'multimodal') {
18993
  await generateMultimodal(messages, id, attData, enableThinking, generationConfig);
18994
  } else {
18995
- await generateCausal(messages, id, generationConfig);
18996
  }
18997
  } else if (type === 'stop') {
18998
  stopping_criteria.interrupt();
 
2674
  <li><strong>Ternary Bonsai 2 27B</strong> (~5.9 GB) &mdash; text + agent. PrismML's Ternary Bonsai 2 27B (Sept 2026; ternary 1.75-bit weights, Qwen3.8-27B backbone, 98% of the full-precision model's benchmark score) on a from-scratch WebGPU engine: hand-written WGSL for ternary dequant + Gated-DeltaNet linear attention, reading the PTQ1_0 GGUF directly. Ported from <em>webml-community/ternary-bonsai-2-webgpu-kernels</em>; Apache-2.0 weights. WebGPU-only &mdash; the biggest in-browser model here. A reasoning model: flip <strong>Show reasoning</strong> in Settings for its chain-of-thought. Text-only here (its optional vision tower is not loaded).</li>
2675
  <li><strong>Underdog Saluki 27B</strong> (~7.9 GB) &mdash; text + agent. ConwayResearch's Underdog Saluki 27B 1.0 (Oct 2026): Qwen3.8-27B as one 2-bit-class IQ2-mix GGUF, tuned to keep tool calling intact (the vendor reports 88 of 120 on a frozen BFCL v4 tool-calling set, against 84 for the full-size model and 70 for Ternary Bonsai 2). Runs on a from-scratch WebGPU engine that reads the GGUF&rsquo;s eleven llama.cpp block types directly; greedy output matches llama.cpp on the same file. The first load copies the 7.9 GB file into this browser&rsquo;s on-device storage once (or pick the file you already have with <strong>Load .gguf from disk&hellip;</strong>); it needs about 8 GB of GPU memory. Apache-2.0 weights. A reasoning model: flip <strong>Show reasoning</strong> in Settings for its chain-of-thought. Text-only here (its optional vision add-on is not loaded).</li>
2676
  <li><strong>Qwen3.5 4B</strong> (~3 GB) &mdash; text + image + agent (tool calling). Qwen3.5 at q4f16, Apache-2.0; its native 262K context is capped to 32K to keep browser memory bounded.</li>
2677
+ <li><strong>LFM3 1.7B</strong> (~1.4 GB) &mdash; text + agent (tool calling). Liquid AI's LFM3-1.7B (Oct 2026), post-trained for agent use, on a new hybrid of short convolutions, sliding-window and global attention. Its tokenizer spends 3 to 4 times fewer tokens on Indic scripts than LFM2.5. A reasoning model: flip <strong>Show reasoning</strong> in Settings for its chain-of-thought. q4f16, 32K context.</li>
2678
  <li><strong>Gemini Nano</strong> (0 MB) &mdash; Google's on-device model built into Chrome, via the Prompt API. Zero download, no WebGPU. Chrome/Edge only; runs where your browser exposes <code>self.LanguageModel</code>. When Nano is already downloaded on desktop Chrome/Edge, LocalMind starts on it by default so you can chat with no model download.</li>
2679
  <li><strong>LFM2.5 230M &middot; WebGPU kernels</strong> (~140 MB) &mdash; the same LFM2.5 230M run on a from-scratch WebGPU engine whose every kernel (RoPE, RMSNorm, Q4_0 dequant, the LFM2 short-conv, GQA attention) is hand-written WGSL, reading a Q4_0 GGUF directly (no ONNX, no llama.cpp). Ported from <em>webml-community/lfm2-webgpu-kernels</em>. WebGPU-only, tuned for maximal decode throughput.</li>
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>
 
2996
  <option value="ternary-bonsai-2-27b-webgpu">Ternary Bonsai 2 27B &middot; Agent (~5.9 GB)</option>
2997
  <option value="underdog-saluki-27b">Underdog Saluki 27B &middot; Agent (~7.9 GB)</option>
2998
  <option value="qwen35-4b">Qwen3.5 4B &middot; Vision + Agent (~3 GB)</option>
2999
+ <option value="lfm3-1.7b">LFM3 1.7B &middot; Agent (~1.4 GB)</option>
3000
  <option value="gemini-nano">Gemini Nano &middot; Built into Chrome (0 MB)</option>
3001
  <option value="gemma4-e2b">Gemma 4 E2B &middot; Multimodal (~1.5 GB)</option>
3002
  <option value="gemma4-e4b">Gemma 4 E4B &middot; Multimodal (~4.9 GB)</option>
 
7483
  // recipe (fused fp16 for WebGPU) are in reference/MDLM-algorithm.md.
7484
 
7485
  import * as ortDefault from 'https://cdn.jsdelivr.net/npm/onnxruntime-web@1.26.0/dist/ort.webgpu.min.mjs';
7486
+ import { AutoTokenizer } from 'https://cdn.jsdelivr.net/npm/@huggingface/transformers@4.3.2/+esm';
7487
 
7488
  // Tiny-A2D Qwen3-0.6B MDLM token ids (tokenizer.mask_token_id / <|im_end|>).
7489
  export const MASK_ID = 151669n;
 
8810
  // WASM by default so it never contends with the chat model's WebGPU. On-device
8811
  // only — the sole network touch is the one-time model fetch. (Non-English turns
8812
  // route to the existing Whisper worker on the main thread instead.)
8813
+ import { pipeline, env } from 'https://cdn.jsdelivr.net/npm/@huggingface/transformers@4.3.2/+esm';
8814
  env.allowLocalModels = false;
8815
  env.allowRemoteModels = true;
8816
 
 
8853
  // every module fp16 (~1.0 GB); Light = q4f16 (~0.36 GB) for phones. (q4f16 was
8854
  // degenerate on the 256M under 4.2.0; the 500M on this 3.7.1 pin is under test.)
8855
  //
8856
+ // VERSION PIN — this vision worker pins transformers.js to 3.7.1, NOT the @4.3.2
8857
  // the chat/OCR workers use. Verified live 2026-08-21: on 4.2.0 SmolVLM's Idefics3
8858
  // processor preprocesses fine but generation is DEGENERATE (pure gibberish); on
8859
  // 3.7.1 the identical model+image+dtype produces a coherent, image-grounded
 
8940
  ChatterboxModel,
8941
  AutoProcessor,
8942
  Tensor,
8943
+ } from 'https://cdn.jsdelivr.net/npm/@huggingface/transformers@4.3.2/+esm';
8944
 
8945
  env.allowLocalModels = true;
8946
  env.localModelPath = '/models/';
 
9055
  'onnx-community/Apertus-v1.1-4B-Instruct-ONNX', 'onnx-community/Qwen3-4B-ONNX',
9056
  'onnx-community/granite-4.0-1b-ONNX-web', 'onnx-community/granite-4.0-micro-ONNX-web',
9057
  'onnx-community/Ternary-Bonsai-4B-ONNX', // first-generation ternary 4B, replaced by Ternary Bonsai 2 4B (2026-10-16)
9058
+ 'LiquidAI/LFM2.5-1.2B-Instruct-ONNX', // replaced by LFM3 1.7B (2026-10-28)
9059
  ];
9060
 
9061
  const MODELS = {
 
9078
  // top_k 20. Apache-2.0, native tool calling.
9079
  genConfig: { temperature: 0.7, top_k: 20, top_p: 0.8, max_new_tokens: 2048, repetition_penalty: 1.05 },
9080
  },
9081
+ 'lfm3-1.7b': {
9082
+ id: 'LiquidAI/LFM3-1.7B-ONNX',
9083
+ label: 'LFM3 1.7B',
9084
  dtype: 'q4f16',
9085
+ size: '~1.4 GB',
9086
  type: 'causal',
9087
  multimodal: false,
9088
  agentCapable: true,
9089
  contextSize: 32768,
9090
+ // Liquid AI LFM3-1.7B (2026-10-23), replacing LFM2.5 1.2B (2026-10-28). model_type 'lfm3' (13 short-conv,
9091
+ // 5 sliding-window and 6 global attention layers), which transformers.js loads from 4.3.2 on. First-party
9092
+ // ONNX (q4f16, 1.38 GB). Post-trained for agent harnesses; its tool calls are the Qwen3-Coder XML form
9093
+ // (<tool_call><function=…><parameter=…>), which parseToolCalls already reads. A reasoning model: the
9094
+ // template thinks unless enable_thinking is false, so generateCausal passes the Show-reasoning toggle.
9095
+ // LFM Open License v1.0. Native 256K context capped to 32K. Liquid's "general" sampling (chat, tools,
9096
+ // agents): temperature 0.1, top_k 50, repetition_penalty 1.1, no top_p.
9097
+ genConfig: { temperature: 0.1, top_k: 50, top_p: 1.0, max_new_tokens: 2048, repetition_penalty: 1.1 },
9098
  },
9099
  'gemini-nano': {
9100
  id: 'chrome-builtin/gemini-nano',
 
10486
  const TOKEN_WEIGHTS = {
10487
  qwen: [0.196, 0.549, 0.473, 0.515, 0.933, 0.717, 1.011, 0.488, 0.644, 0.612, 0.615, 0.291, 0.236, 0.371, 2.845],
10488
  lfm2: [0.202, 0.578, 1.495, 2.353, 2.97, 2.964, 2.971, 2.494, 2.966, 2.956, 2.98, 0.485, 0.357, 0.446, 5.534],
10489
+ lfm3: [0.218, 0.387, 0.578, 0.682, 0.851, 0.661, 0.858, 0.716, 0.682, 0.66, 0.643, 0.312, 0.334, 0.494, 2.457],
10490
  gemma: [0.199, 0.618, 0.28, 0.237, 0.659, 0.424, 0.908, 0.25, 0.361, 0.34, 0.331, 0.352, 0.233, 0.412, 2.692],
10491
  minicpm5: [0.212, 0.427, 0.902, 1.846, 3.18, 3.191, 3.124, 1.986, 2.127, 1.452, 2.386, 1.054, 0.278, 0.626, 0.01],
10492
  unknown: [0.244, 0.711, 1.719, 2.706, 3.657, 3.67, 3.593, 2.868, 3.411, 3.4, 3.427, 1.212, 0.411, 0.729, 6.364],
 
10504
  'gemma4-26b-a4b-ssd': 'gemma',
10505
  'ternary-bonsai-2-2b-webgpu': 'qwen',
10506
  'ternary-bonsai-2-4b-webgpu': 'qwen',
10507
+ 'lfm3-1.7b': 'lfm3',
10508
  'lfm2-230m-gguf': 'lfm2',
10509
  'lfm2-230m-webgpu': 'lfm2',
10510
  'gemma4-e2b': 'gemma',
 
17622
  pipeline,
17623
  AutoTokenizer,
17624
  AutoModelForSequenceClassification,
17625
+ } from 'https://cdn.jsdelivr.net/npm/@huggingface/transformers@4.3.2/+esm';
17626
 
17627
  env.allowLocalModels = true;
17628
  env.localModelPath = '/models/';
 
17933
  Sam3TrackerModel,
17934
  AutoProcessor,
17935
  RawImage,
17936
+ } from 'https://cdn.jsdelivr.net/npm/@huggingface/transformers@4.3.2/+esm';
17937
 
17938
  env.allowLocalModels = true;
17939
  env.localModelPath = '/models/';
 
18074
  RawImage,
18075
  TextStreamer,
18076
  InterruptableStoppingCriteria,
18077
+ } from 'https://cdn.jsdelivr.net/npm/@huggingface/transformers@4.3.2/+esm';
18078
 
18079
  env.allowLocalModels = true;
18080
  env.localModelPath = '/models/';
 
18247
  AutoTokenizer,
18248
  WhisperForConditionalGeneration,
18249
  full,
18250
+ } from 'https://cdn.jsdelivr.net/npm/@huggingface/transformers@4.3.2/+esm';
18251
 
18252
  env.allowLocalModels = true;
18253
  env.localModelPath = '/models/';
 
18726
  // would race the 'load' postMessage from the main thread — it can arrive
18727
  // before the module-body handler is registered and get dropped. Kick off
18728
  // the import eagerly and have the message handler await it.
18729
+ let env, AutoTokenizer, AutoModelForCausalLM, AutoProcessor, Gemma4ForConditionalGeneration, Qwen3_5ForConditionalGeneration, load_image, TextStreamer, InterruptableStoppingCriteria, Tensor;
18730
  let stopping_criteria;
18731
 
18732
+ const __modReady = import('https://cdn.jsdelivr.net/npm/@huggingface/transformers@4.3.2/+esm').then((m) => {
18733
  env = m.env;
18734
  AutoTokenizer = m.AutoTokenizer;
18735
  AutoModelForCausalLM = m.AutoModelForCausalLM;
 
18739
  load_image = m.load_image;
18740
  TextStreamer = m.TextStreamer;
18741
  InterruptableStoppingCriteria = m.InterruptableStoppingCriteria;
18742
+ Tensor = m.Tensor;
18743
  env.allowLocalModels = true;
18744
  env.localModelPath = '/models/';
18745
  env.allowRemoteModels = true;
 
18759
  self.postMessage({ type: 'progress', data: p });
18760
  }
18761
 
18762
+ // LiquidAI's LFM3 ONNX export names its last-position input logits_to_keep (shape [1]); transformers.js 4.3.2 feeds
18763
+ // only num_logits_to_keep and rejects the run as "Missing the following inputs: logits_to_keep". Hide the input from
18764
+ // transformers.js's check and add it at the session's run, set to 1: generate() reads only the last position's logits.
18765
+ // Checked in Chrome on this Mac, 2026-10-28: a 20-token and a 1,176-token prompt both answer correctly.
18766
+ function feedLogitsToKeep(m) {
18767
+ const sess = m && m.sessions && m.sessions.model;
18768
+ if (!sess || !sess.inputNames.includes('logits_to_keep')) return;
18769
+ const names = sess.inputNames.filter((n) => n !== 'logits_to_keep');
18770
+ const one = new Tensor('int64', [1n], [1]).ort_tensor;
18771
+ const run = sess.run.bind(sess);
18772
+ Object.defineProperty(sess, 'inputNames', { value: names });
18773
+ sess.run = (feeds, opts) => run({ ...feeds, logits_to_keep: one }, opts);
18774
+ }
18775
+
18776
  async function loadCausal(modelId, dtype) {
18777
  processor = null; tokenizer = null; model = null;
18778
 
 
18785
  device: 'webgpu',
18786
  progress_callback: progressCallback,
18787
  });
18788
+ feedLogitsToKeep(model);
18789
 
18790
  // Warmup: compile WebGPU shaders with a single token
18791
  self.postMessage({ type: 'warmup' });
 
18808
  ? Qwen3_5ForConditionalGeneration
18809
  : Gemma4ForConditionalGeneration;
18810
  if (!ConditionalGenerationModel) {
18811
+ throw new Error('Transformers.js 4.3.2 does not export ' + modelClass);
18812
  }
18813
  model = await ConditionalGenerationModel.from_pretrained(modelId, {
18814
  dtype: dtype,
 
18825
  self.postMessage({ type: 'ready' });
18826
  }
18827
 
18828
+ async function generateCausal(chatMessages, id, genConfig, enableThinking) {
18829
  stopping_criteria.reset();
18830
+ // Templates that think by default (LFM3) read enable_thinking; the others ignore it.
18831
+ const thinking = enableThinking === true;
18832
 
18833
  try {
18834
  let inputs;
 
18836
  inputs = tokenizer.apply_chat_template(chatMessages, {
18837
  add_generation_prompt: true,
18838
  return_dict: true,
18839
+ enable_thinking: thinking,
18840
  });
18841
  } catch (tplErr) {
18842
  // The pinned transformers.js jinja can't parse some real templates — e.g.
 
18859
  inputs = tokenizer.apply_chat_template(chatMessages, {
18860
  add_generation_prompt: true,
18861
  return_dict: true,
18862
+ enable_thinking: thinking,
18863
  chat_template: raw,
18864
  });
18865
  salvaged = true;
 
18882
  }
18883
  }
18884
 
18885
+ // A thinking template ends the prompt on an open <think> (LFM3), so the reply starts inside the thought: open
18886
+ // LocalMind's <|think|> marker first and swap the model's </think> for <|/think|>. With thinking off LFM3 still
18887
+ // opens <think> itself on some prompts (Liquid's card says so; seen on a 1,176-token prompt), so that becomes
18888
+ // <|think|> too. Both tags are one token there, so each arrives whole in one streamer chunk.
18889
+ const promptText = thinking ? tokenizer.decode(inputs.input_ids.tolist()[0].slice(-4).map(Number)) : '';
18890
+ let inThink = thinking && promptText.trimEnd().endsWith('<think>');
18891
+ if (inThink) self.postMessage({ type: 'token', token: '<|think|>', id });
18892
  const streamer = new TextStreamer(tokenizer, {
18893
  skip_prompt: true,
18894
  skip_special_tokens: true,
18895
  callback_function: (text) => {
18896
+ if (!inThink && text.indexOf('<think>') !== -1) {
18897
+ text = text.split('<think>').join('<|think|>');
18898
+ inThink = true;
18899
+ }
18900
+ if (inThink && text.indexOf('</think>') !== -1) {
18901
+ text = text.split('</think>').join('<|/think|>');
18902
+ inThink = false;
18903
+ }
18904
  self.postMessage({ type: 'token', token: text, id });
18905
  },
18906
  });
 
19029
  if (loadedType === 'multimodal') {
19030
  await generateMultimodal(messages, id, attData, enableThinking, generationConfig);
19031
  } else {
19032
+ await generateCausal(messages, id, generationConfig, enableThinking);
19033
  }
19034
  } else if (type === 'stop') {
19035
  stopping_criteria.interrupt();