Views
No views yet
cognitivecomputations/dolphin-2.2.1-mistral-7b for @mlc-ai/web-llm ≥ 0.2.82 (i.e. MLC modelVersion v0_2_80), so you can run Dolphin in the browser again in 2026.hrishioa/mlc-chat-dolphin-2.2.1-mistral-7b-q4f32_1 was compiled against an older MLC/TVM runtime and no longer instantiates on current web-llm — you get:LinkError: WebAssembly.instantiate():
Import #2 "env" "TVMWasmPackedCFunc": function import requires a callablemlc_llm convert_weight --quantization q4f32_1), regenerating mlc-chat-config.json, and reusing the prebuilt Mistral-7B architecture WebGPU kernel library from mlc-ai/binary-mlc-llm-libs (since Dolphin 2.2.1 is a Mistral-7B fine-tune, the kernel is identical; only the weights change).params_shard_*.bin — 107 shards, q4f32_1 group quantization, ~3.8 GB total, 5.001 bits/param, 7.24 B parameters.ndarray-cache.json / tensor-cache.json / tensor-cache-b16.json — MLC weight index.mlc-chat-config.json — Mistral-7B config with ChatML conv template, stop_token_ids: [2, 32000], context 4096, prefill chunk 1024, sliding window disabled.tokenizer.json — borrowed from mlc-ai/OpenHermes-2.5-Mistral-7B-q4f16_1-MLC, which uses the exact same Mistral-7B SentencePiece + ChatML added tokens (<|im_end|>=32000, <|im_start|>=32001 with "special": true).tokenizer.model, tokenizer_config.json, added_tokens.json — original SentencePiece files from the Dolphin release, kept as fallbacks.1import { CreateMLCEngine, prebuiltAppConfig } from "@mlc-ai/web-llm";
2
3const modelId = "Dolphin-2.2.1-Mistral-7B-q4f32_1-MLC";
4const engine = await CreateMLCEngine(modelId, {
5 appConfig: {
6 model_list: [
7 ...prebuiltAppConfig.model_list,
8 {
9 model_id: modelId,
10 model: "https://huggingface.co/hrishioa/Dolphin-2.2.1-Mistral-7B-q4f32_1-MLC",
11 model_lib:
12 "https://raw.githubusercontent.com/mlc-ai/binary-mlc-llm-libs/main/web-llm-models/v0_2_80/Mistral-7B-Instruct-v0.3-q4f32_1-ctx4k_cs1k-webgpu.wasm",
13 },
14 ],
15 useIndexedDBCache: prebuiltAppConfig.useIndexedDBCache,
16 },
17});
18
19const reply = await engine.chat.completions.create({
20 messages: [{ role: "user", content: "Hello!" }],
21});wasm-ai project's resurrection_log.md (Attempt 5 and 5b).cognitivecomputations/dolphin-2.2.1-mistral-7b.mistralai/Mistral-7B-v0.1.hrishioa, Apr 2026.