NakliTechie commited on
Commit
f8a3477
·
1 Parent(s): 3e0be79

Sync from GitHub 2026-10-31T07:08:57Z

Browse files
Files changed (1) hide show
  1. index.html +31 -1
index.html CHANGED
@@ -2694,6 +2694,7 @@
2694
  <li><strong>Ling 3.0 Flash 124B-A5B &middot; via anyformer (experimental)</strong> (~81 GB) &mdash; text, reasoning. inclusionAI's Ling 3.0 Flash (released 2026-08-02, MIT): a 124B mixture-of-experts (512 experts, 8 active per token, about 5.1B parameters per token) whose layers mix linear attention (KDA) and latent attention (MLA), from inclusionAI's own 4-bit Q4_K_M GGUF (77 GB in two parts). anyformer loads it on diskformer's engine for it, which keeps the attention, routers and caches on the GPU (about 2.5 GB plus a 4 GB expert cache; it needs about 7 GB of GPU memory) and streams the 20,480 routed experts from on-device storage (OPFS). The first load copies the model into this browser's storage once (~81 GB of disk). About 5.5 tokens/s on an M4 Pro with 24 GB, and prompts are read at about 5 tokens/s. Flip <strong>Show reasoning</strong> for its thinking. Tools are off for now.</li>
2695
  <li><strong>DeepSeek V4.1 Flash 763B &middot; tech demo, very slow, chat untested</strong> (~231 GB) &mdash; text. DeepSeek's 763B mixture-of-experts model, 6 of its 384 experts per token, from ISTA-DASLab's IQ1_M files plus a separate Engram file. anyformer runs it on diskformer's llama.cpp path: an 8 GB GPU budget, with the experts read from this browser's storage as each token needs them. It needs 231 GB of free browser storage and runs at under one token a second, so treat it as a demonstration, not a chat model. One plain prompt was checked in Chrome; chat through its template has not been run, so replies may come out wrong. Reasoning and tools are off.</li>
2696
  <li><strong>Nemotron-Labs-Diffusion 3B &middot; experimental &middot; diffusion</strong> (~2.7 GB) &mdash; text. NVIDIA's Nemotron-Labs-Diffusion-3B writes by diffusion: it fills in whole blocks of words at once instead of one word at a time. It runs on kohra.js through anyformer, at about 12-16 tokens a second on an M4 Pro. Opt-in and experimental: only single-turn chats have been checked. Reasoning and tools are off.</li>
 
2697
  <li><strong>Clover 35B-A3B &middot; SSD-streamed, experimental &middot; diffusion</strong> (~19.6 GB) &mdash; text. Diffuspace's Clover-35B-A3B-Preview is Qwen3.6 35B-A3B retrained to write by diffusion: each block of 32 words is filled in over several passes, so replies appear a block at a time. It streams its experts from on-device storage through anyformer and diskformer, using about 4 GB of GPU memory. The first load copies a 19.6 GB file into browser storage. Opt-in and experimental: its speed has not been measured yet. Tools are off.</li>
2698
  <li><strong>Nemotron 3.5 Bolt E4B &middot; via anyformer (experimental)</strong> (~4.9 GB) &mdash; text. NVIDIA's Nemotron 3.5 Bolt E4B, a dense model with about 4B effective parameters, post-trained from Google's Gemma 4 E4B. It runs on Transformers.js through anyformer, from a community 4-bit ONNX conversion (text only), at about 35 tokens a second on an M4 Pro. Reasoning and tools are off for now.</li>
2699
  <li><strong>GLM-4.7-Flash 30B-A3B &middot; via anyformer (experimental)</strong> (~18 GB) &mdash; text, reasoning. Z.ai's GLM-4.7-Flash, a 30B mixture-of-experts with about 3B parameters per token, from unsloth's 4-bit GGUF (18.3 GB). LocalMind has no engine of its own for it: anyformer picks diskformer's llama.cpp path, which runs llama.cpp's WebGPU build in a worker and pages the experts in from on-device storage (OPFS). The first load copies the file into this browser's storage once. Flip <strong>Show reasoning</strong> for its thinking. Tools are off for now.</li>
@@ -3023,6 +3024,7 @@
3023
  <option value="minicpm-v47-35b-a3b-ssd">MiniCPM-V 4.7 35B-A3B &middot; SSD-streamed, experimental, text and images (~20 GB)</option>
3024
  <option value="gemma4-26b-a4b-ssd">Gemma 4 26B-A4B &middot; SSD-streamed, experimental (~14 GB)</option>
3025
  <option value="lfm25-24b-a2b-ssd">LFM2.5 24B-A2B &middot; SSD-streamed, experimental (~13.6 GB)</option>
 
3026
  <option value="flint-40b-a2b-ssd">Flint 40B-A2.5B &middot; SSD-streamed, experimental, coding (~26 GB)</option>
3027
  <option value="ling-3-flash-anyformer">Ling 3.0 Flash 124B-A5B &middot; via anyformer, experimental (~81 GB)</option>
3028
  <option value="nemotron-bolt-e4b-anyformer">Nemotron 3.5 Bolt E4B &middot; via anyformer, experimental (~4.9 GB)</option>
@@ -9519,6 +9521,26 @@ self.addEventListener('message', async (e) => {
9519
  contextSize: 4096,
9520
  genConfig: { max_new_tokens: 2048 },
9521
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9522
  // backend:'flint-ssd' → the custom-WGSL FlintMoeSsd engine (flint_moe_ssd.js, on the LFM2.5 rung): Recursive's Flint
9523
  // (2026-10-21, a 39.8B coding MoE, ~2.5B active) from Recursive's own Q4_1 GGUF (26 GB, not gated; the card asks for
9524
  // 48 GB of memory under llama.cpp). Attention, routers, the F32 token table and KV stay on the GPU; the 12,032
@@ -17747,6 +17769,7 @@ self.addEventListener('message', async (e) => {
17747
  'anyformer-gemma26b': { engine: 'anyformer_models.js', exportName: 'Gemma4_26B', label: 'Gemma 4 26B-A4B (Q4_0, via anyformer)' },
17748
  'anyformer-flint': { engine: 'anyformer_models.js', exportName: 'Flint40B', label: 'Flint 40B-A2.5B (Q4_1, via anyformer)' },
17749
  'anyformer-lfm25-24b': { engine: 'anyformer_models.js', exportName: 'Lfm25_24B', label: 'LFM2.5 24B-A2B (Q4_0, via anyformer)' },
 
17750
  'anyformer-qwen36': { engine: 'anyformer_models.js', exportName: 'Qwen36', label: 'Qwen3.6 35B-A3B (Q8_0, via anyformer)' },
17751
  'anyformer-ling3flash': { engine: 'anyformer_models.js', exportName: 'Ling3Flash', label: 'Ling 3.0 Flash (Q4_K_M, via anyformer)' },
17752
  'anyformer-glm47flash': { engine: 'anyformer_models.js', exportName: 'Glm47Flash', label: 'GLM-4.7-Flash (Q4_K_M, via anyformer)' },
@@ -19429,7 +19452,7 @@ self.addEventListener('message', async (e) => {
19429
  const ENGINE_STORES = {
19430
  'transformers': ['transformers-cache'], 'wllama': ['wllama'], 'lfm2-webgpu': ['gguf-v1'],
19431
  'bonsai2-webgpu': ['gguf-cache-v1'], 'gemma4-webgpu': ['safetensors-cache-v1', 'bonsai-pipeline-v1'],
19432
- 'qwen35moe-ssd': ['on disk (OPFS)'], 'minicpmv-ssd': ['on disk (OPFS)'], 'gemma4moe-ssd': ['on disk (OPFS)'], 'lfm2moe-ssd': ['on disk (OPFS)'], 'flint-ssd': ['on disk (OPFS)'], 'anyformer-glm47flash': ['on disk (OPFS)'], 'anyformer-deepseek-v41': ['on disk (OPFS)'], 'anyformer-bolt-e4b': ['transformers-cache'], 'anyformer-lfm3-omni': ['diskformer-models'], 'anyformer-nemotron-diffusion': ['kohra-models', 'transformers-cache'], 'anyformer-ling3flash': ['on disk (OPFS)'], 'anyformer-qwen36': ['on disk (OPFS)'], 'anyformer-clover': ['on disk (OPFS)'], 'anyformer-flint': ['on disk (OPFS)'], 'anyformer-lfm25-24b': ['on disk (OPFS)'], 'anyformer-gemma26b': ['on disk (OPFS)'], 'qwen3dense': ['on disk (OPFS)'], 'qwen35dense': ['on disk (OPFS)'],
19433
  };
19434
  function parseSizeBytes(size) {
19435
  const m = /([\d.]+)\s*(GB|MB)/i.exec(String(size || ''));
@@ -24373,6 +24396,13 @@ export class Clover35B extends AnyformerModel {
24373
  static entry = 'CLOVER_35B_A3B';
24374
  }
24375
 
 
 
 
 
 
 
 
24376
  export class DeepSeekV41Flash extends AnyformerModel {
24377
  static entry = 'DEEPSEEK_V4_1_FLASH';
24378
  }
 
2694
  <li><strong>Ling 3.0 Flash 124B-A5B &middot; via anyformer (experimental)</strong> (~81 GB) &mdash; text, reasoning. inclusionAI's Ling 3.0 Flash (released 2026-08-02, MIT): a 124B mixture-of-experts (512 experts, 8 active per token, about 5.1B parameters per token) whose layers mix linear attention (KDA) and latent attention (MLA), from inclusionAI's own 4-bit Q4_K_M GGUF (77 GB in two parts). anyformer loads it on diskformer's engine for it, which keeps the attention, routers and caches on the GPU (about 2.5 GB plus a 4 GB expert cache; it needs about 7 GB of GPU memory) and streams the 20,480 routed experts from on-device storage (OPFS). The first load copies the model into this browser's storage once (~81 GB of disk). About 5.5 tokens/s on an M4 Pro with 24 GB, and prompts are read at about 5 tokens/s. Flip <strong>Show reasoning</strong> for its thinking. Tools are off for now.</li>
2695
  <li><strong>DeepSeek V4.1 Flash 763B &middot; tech demo, very slow, chat untested</strong> (~231 GB) &mdash; text. DeepSeek's 763B mixture-of-experts model, 6 of its 384 experts per token, from ISTA-DASLab's IQ1_M files plus a separate Engram file. anyformer runs it on diskformer's llama.cpp path: an 8 GB GPU budget, with the experts read from this browser's storage as each token needs them. It needs 231 GB of free browser storage and runs at under one token a second, so treat it as a demonstration, not a chat model. One plain prompt was checked in Chrome; chat through its template has not been run, so replies may come out wrong. Reasoning and tools are off.</li>
2696
  <li><strong>Nemotron-Labs-Diffusion 3B &middot; experimental &middot; diffusion</strong> (~2.7 GB) &mdash; text. NVIDIA's Nemotron-Labs-Diffusion-3B writes by diffusion: it fills in whole blocks of words at once instead of one word at a time. It runs on kohra.js through anyformer, at about 12-16 tokens a second on an M4 Pro. Opt-in and experimental: only single-turn chats have been checked. Reasoning and tools are off.</li>
2697
+ <li><strong>LFM2.5 24B-A2B + Pi coding adapter &middot; experimental</strong> (~13.6 GB) &mdash; text. The LFM2.5 24B-A2B model above with harness-hub's Pi coding-agent adapter (37 MB) applied when it loads. The adapter was trained inside the Pi coding agent, so outside it expect behaviour like the base model or worse. It runs on llama.cpp in the tab with experts read from on-device storage, about 4 GB of GPU memory, and is slower than the base entry (about 3 tokens a second on an M4 Pro). Tools are off. Licence lfm1.0: royalty-free commercial use only below US$10M annual revenue.</li>
2698
  <li><strong>Clover 35B-A3B &middot; SSD-streamed, experimental &middot; diffusion</strong> (~19.6 GB) &mdash; text. Diffuspace's Clover-35B-A3B-Preview is Qwen3.6 35B-A3B retrained to write by diffusion: each block of 32 words is filled in over several passes, so replies appear a block at a time. It streams its experts from on-device storage through anyformer and diskformer, using about 4 GB of GPU memory. The first load copies a 19.6 GB file into browser storage. Opt-in and experimental: its speed has not been measured yet. Tools are off.</li>
2699
  <li><strong>Nemotron 3.5 Bolt E4B &middot; via anyformer (experimental)</strong> (~4.9 GB) &mdash; text. NVIDIA's Nemotron 3.5 Bolt E4B, a dense model with about 4B effective parameters, post-trained from Google's Gemma 4 E4B. It runs on Transformers.js through anyformer, from a community 4-bit ONNX conversion (text only), at about 35 tokens a second on an M4 Pro. Reasoning and tools are off for now.</li>
2700
  <li><strong>GLM-4.7-Flash 30B-A3B &middot; via anyformer (experimental)</strong> (~18 GB) &mdash; text, reasoning. Z.ai's GLM-4.7-Flash, a 30B mixture-of-experts with about 3B parameters per token, from unsloth's 4-bit GGUF (18.3 GB). LocalMind has no engine of its own for it: anyformer picks diskformer's llama.cpp path, which runs llama.cpp's WebGPU build in a worker and pages the experts in from on-device storage (OPFS). The first load copies the file into this browser's storage once. Flip <strong>Show reasoning</strong> for its thinking. Tools are off for now.</li>
 
3024
  <option value="minicpm-v47-35b-a3b-ssd">MiniCPM-V 4.7 35B-A3B &middot; SSD-streamed, experimental, text and images (~20 GB)</option>
3025
  <option value="gemma4-26b-a4b-ssd">Gemma 4 26B-A4B &middot; SSD-streamed, experimental (~14 GB)</option>
3026
  <option value="lfm25-24b-a2b-ssd">LFM2.5 24B-A2B &middot; SSD-streamed, experimental (~13.6 GB)</option>
3027
+ <option value="lfm25-24b-pi-anyformer">LFM2.5 24B-A2B + Pi coding adapter &middot; experimental (~13.6 GB)</option>
3028
  <option value="flint-40b-a2b-ssd">Flint 40B-A2.5B &middot; SSD-streamed, experimental, coding (~26 GB)</option>
3029
  <option value="ling-3-flash-anyformer">Ling 3.0 Flash 124B-A5B &middot; via anyformer, experimental (~81 GB)</option>
3030
  <option value="nemotron-bolt-e4b-anyformer">Nemotron 3.5 Bolt E4B &middot; via anyformer, experimental (~4.9 GB)</option>
 
9521
  contextSize: 4096,
9522
  genConfig: { max_new_tokens: 2048 },
9523
  },
9524
+ // backend:'anyformer-lfm25-24b-pi' → anyformer's MODELS.chat.LFM2_5_24B_PI (0.4.18+): the same QAD-Q4_0 GGUF with
9525
+ // harness-hub's Pi coding-agent LoRA (37.4 MB, applied at load, not merged) on diskformer's llama.cpp path with paged
9526
+ // experts, a 4 GB GPU budget. anyformer 0.4.19 (diskformer 0.6.18) closes the template's always-open <think> when
9527
+ // thinking is off. anyformer's check: 4/4 turns on topic, 2.6-3.1 tok/s, Chrome 156 on an M4 Pro. Shares the base
9528
+ // entry's id, so resolveModelKey('LiquidAI/LFM2.5-24B-A2B-GGUF') still picks the base; this one is reached by key.
9529
+ // Tools off until LocalMind's tool battery runs on it (the adapter was trained inside the Pi agent's own harness).
9530
+ // Licence lfm1.0 (model and adapter).
9531
+ 'lfm25-24b-pi-anyformer': {
9532
+ id: 'LiquidAI/LFM2.5-24B-A2B-GGUF',
9533
+ label: 'LFM2.5 24B-A2B + Pi coding adapter · experimental',
9534
+ backend: 'anyformer-lfm25-24b-pi',
9535
+ size: '~13.6 GB',
9536
+ gguf: { file: 'LFM2.5-24B-A2B-QAD-Q4_0.gguf', size: 13578190880 },
9537
+ type: 'causal',
9538
+ multimodal: false,
9539
+ agentCapable: false,
9540
+ thinking: true,
9541
+ contextSize: 4096,
9542
+ genConfig: { max_new_tokens: 2048 },
9543
+ },
9544
  // backend:'flint-ssd' → the custom-WGSL FlintMoeSsd engine (flint_moe_ssd.js, on the LFM2.5 rung): Recursive's Flint
9545
  // (2026-10-21, a 39.8B coding MoE, ~2.5B active) from Recursive's own Q4_1 GGUF (26 GB, not gated; the card asks for
9546
  // 48 GB of memory under llama.cpp). Attention, routers, the F32 token table and KV stay on the GPU; the 12,032
 
17769
  'anyformer-gemma26b': { engine: 'anyformer_models.js', exportName: 'Gemma4_26B', label: 'Gemma 4 26B-A4B (Q4_0, via anyformer)' },
17770
  'anyformer-flint': { engine: 'anyformer_models.js', exportName: 'Flint40B', label: 'Flint 40B-A2.5B (Q4_1, via anyformer)' },
17771
  'anyformer-lfm25-24b': { engine: 'anyformer_models.js', exportName: 'Lfm25_24B', label: 'LFM2.5 24B-A2B (Q4_0, via anyformer)' },
17772
+ 'anyformer-lfm25-24b-pi': { engine: 'anyformer_models.js', exportName: 'Lfm25_24bPi', label: 'LFM2.5 24B-A2B + Pi LoRA (Q4_0, via anyformer and diskformer llama.cpp)' },
17773
  'anyformer-qwen36': { engine: 'anyformer_models.js', exportName: 'Qwen36', label: 'Qwen3.6 35B-A3B (Q8_0, via anyformer)' },
17774
  'anyformer-ling3flash': { engine: 'anyformer_models.js', exportName: 'Ling3Flash', label: 'Ling 3.0 Flash (Q4_K_M, via anyformer)' },
17775
  'anyformer-glm47flash': { engine: 'anyformer_models.js', exportName: 'Glm47Flash', label: 'GLM-4.7-Flash (Q4_K_M, via anyformer)' },
 
19452
  const ENGINE_STORES = {
19453
  'transformers': ['transformers-cache'], 'wllama': ['wllama'], 'lfm2-webgpu': ['gguf-v1'],
19454
  'bonsai2-webgpu': ['gguf-cache-v1'], 'gemma4-webgpu': ['safetensors-cache-v1', 'bonsai-pipeline-v1'],
19455
+ 'qwen35moe-ssd': ['on disk (OPFS)'], 'minicpmv-ssd': ['on disk (OPFS)'], 'gemma4moe-ssd': ['on disk (OPFS)'], 'lfm2moe-ssd': ['on disk (OPFS)'], 'flint-ssd': ['on disk (OPFS)'], 'anyformer-glm47flash': ['on disk (OPFS)'], 'anyformer-deepseek-v41': ['on disk (OPFS)'], 'anyformer-bolt-e4b': ['transformers-cache'], 'anyformer-lfm3-omni': ['diskformer-models'], 'anyformer-nemotron-diffusion': ['kohra-models', 'transformers-cache'], 'anyformer-ling3flash': ['on disk (OPFS)'], 'anyformer-qwen36': ['on disk (OPFS)'], 'anyformer-clover': ['on disk (OPFS)'], 'anyformer-flint': ['on disk (OPFS)'], 'anyformer-lfm25-24b': ['on disk (OPFS)'], 'anyformer-lfm25-24b-pi': ['on disk (OPFS)'], 'anyformer-gemma26b': ['on disk (OPFS)'], 'qwen3dense': ['on disk (OPFS)'], 'qwen35dense': ['on disk (OPFS)'],
19456
  };
19457
  function parseSizeBytes(size) {
19458
  const m = /([\d.]+)\s*(GB|MB)/i.exec(String(size || ''));
 
24396
  static entry = 'CLOVER_35B_A3B';
24397
  }
24398
 
24399
+ // LFM2.5 24B-A2B with harness-hub's Pi coding-agent LoRA by anyformer's catalog entry (0.4.18+; 0.4.19 closes the
24400
+ // always-open <think> when thinking is off): the QAD-Q4_0 GGUF plus a 37.4 MB adapter applied at load, on diskformer's
24401
+ // llama.cpp path with paged experts, a 4 GB budget. Opt-in and experimental; licence lfm1.0.
24402
+ export class Lfm25_24bPi extends AnyformerModel {
24403
+ static entry = 'LFM2_5_24B_PI';
24404
+ }
24405
+
24406
  export class DeepSeekV41Flash extends AnyformerModel {
24407
  static entry = 'DEEPSEEK_V4_1_FLASH';
24408
  }