Modelpublic

skillsafe-ai/qwen3-0.6b-q4f16

sourceHugging Faceapache-2.0updated 5d agoView on Hugging Face

0likes

Model Card

Qwen3-0.6B (q4f16_1, WebLLM)

Browser-ready mlc artifacts for text-generation, produced by SkillSafe's reproducible converter (`models/` in skillsafe.ai) from a pinned upstream source. Every byte here is derivable from that source plus the recipe below; nothing was edited by hand.

Provenance


Upstream	https://huggingface.co/mlc-ai/Qwen3-0.6B-q4f16_1-MLC/tree/8c14ce481d4c692769976ad52afea453a102df19
Upstream SHA-256 / commit	`8c14ce481d4c692769976ad52afea453a102df19`
Recipe	`recipes/qwen3-0.6b-q4f16.yaml` — sha256 `94784f18373a8083b0a13184727c0dd088f2310ae8f7635fec1e42e139669df9`
Toolchain	Python 3.12.13, torch 2.10.0, onnx 1.23.0, onnxruntime 1.30.0 on Darwin 25.6.0 arm64
Converted	2026-09-22T19:00:49+00:00

Files

file	class	size	SHA-256
`lib/Qwen3-0.6B-q4f16_1-ctx4k_cs1k-webgpu.wasm`	registry-shared	5.49 MB	`7a51cff80ad5fe1539e033a83b9a7d240957d37e3d8e620c6b43648aba90a0cc`
`merges.txt`	bundle	1.59 MB	`8831e4f1a044471340f7c0a83d7bd71306a5b867e95fd870f74d0c5308a904d5`
`mlc-chat-config.json`	bundle	0.00 MB	`190a51c56b9a07a8d8721266f8e459bcd1b1689bcdff294795cd2bffa203ff2d`
`ndarray-cache.json`	bundle	0.12 MB	`f4647d617dc869d91fbd7851c65af0a2cf7dd615884ebe3b110013801b65d8c7`
`params_shard_0.bin`	registry	74.19 MB	`61062e64b475309e648516c853c7bd949344f1ce5c0921e9b121f905953ce9e3`
`params_shard_1.bin`	registry	31.22 MB	`a7d9e88e510636b34c850bf19c660c38cdacbe64bd128c6bb4dcc710455db81d`
`params_shard_2.bin`	registry	30.39 MB	`eb5af136518c33abd7c2eb66e154b5504ce27f62728aafffd8fca26a3cab64a9`
`params_shard_3.bin`	registry	31.95 MB	`482d32bfefe36ec9ee452d88c25cce42f31d8350e4abccc8f4727e4006ab2859`
`params_shard_4.bin`	registry	30.52 MB	`b3d1ecb2553ae323c95b19e231dc74e1ef34d63094ccf0dd3e79a37f312663f3`
`params_shard_5.bin`	registry	30.39 MB	`2c753a811fbea6e0292caf39d9f9ceeb8cb5ff74a183f7350be3f846f3013438`
`params_shard_6.bin`	registry	31.95 MB	`52d543cee1e1d6526dae4a5a36a69c1bafa222d1e3e95c4ee46456d56f1aeab2`
`params_shard_7.bin`	registry	30.52 MB	`3d9047c2603d24b4ca8e689bc924a7715224205ba591504238ff5910e8542c65`
`params_shard_8.bin`	registry	28.70 MB	`13985d7899bf0188ff5a0d37fe6254a8b3e3bf7e4aba0d198de1469faa9ef9b4`
`tensor-cache.json`	bundle	0.12 MB	`f4647d617dc869d91fbd7851c65af0a2cf7dd615884ebe3b110013801b65d8c7`
`tokenizer.json`	bundle	10.89 MB	`aeb13307a71acd8fe81861d94ad54ab689df773318809eed3cbe794b4492dae4`
`tokenizer_config.json`	bundle	0.01 MB	`bbc2c089e3bef8753f63348c10595767be8998bc6ca7c7ebcaada346d401fba8`
`vocab.json`	bundle	2.65 MB	`ca10d7e9fb3ed18575dd1e277a2579c16d108e32f27439684afa0e10b1440910`

registry files are parameter files served from models.skillsafe.ai once vetted; bundle files ship inside an app; registry-shared is a runtime library reused by every model of the same architecture.

Verification

Shard sizes and md5s match ndarray-cache.json; the implied parameter count matches the checkpoint (751,632,384 parameters, 3.57 bits/param); the parameter table embedded in the WebGPU model library matches all 339 tensors of these weights.

Use in the browser (WebLLM)

import { CreateMLCEngine } from "@mlc-ai/web-llm";
const engine = await CreateMLCEngine("qwen3-0.6b-q4f16", {
  appConfig: { model_list: [{
    model: "https://huggingface.co/skillsafe-ai/qwen3-0.6b-q4f16",
    model_id: "qwen3-0.6b-q4f16",
    model_lib: "https://huggingface.co/skillsafe-ai/qwen3-0.6b-q4f16/resolve/main/lib/Qwen3-0.6B-q4f16_1-ctx4k_cs1k-webgpu.wasm",
    overrides: { context_window_size: 4096, prefill_chunk_size: 1024 },
  }] },
});

The model library is compiled for context window 4096 / prefill chunk 1024; mlc-chat-config.json says 40960 / 2048, so set the overrides above.

Licence and attribution

Licence: Apache-2.0 — notice: https://huggingface.co/Qwen/Qwen3-0.6B/blob/c1899de289a04d12100db370d81485cdf75e47ca/LICENSE. The conversion recipe and this model card are part of the SkillSafe repository and carry its licence; the weights remain under the upstream licence above.

The full manifest.json in this repo records the recipe, sources, toolchain (including the uv.lock hash) and per-file verification numbers.