--- license: apache-2.0 base_model: Qwen/Qwen3-8B tags: - quantization - trellis-coded-quantization - qtip - webgpu - 3-bit - on-device library_name: transformers --- # Qwen3-8B — 3-bit trellis-quantized for WebGPU A **3-bit trellis-coded quantization (TCQ)** of [Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B), packed to run its **full forward pass in a web browser on WebGPU** — no CUDA, no server. Runs on a **4 GB GPU**: verified live in Chrome on a laptop RTX 3050 Ti, where it completed *"The capital of France is"* → *"Paris. The capital of Italy is Rome. The capital of Germany is Berlin."* This is part of **[Trellis WebGPU](https://github.com/Azimml/trellis-webgpu)** — the first trellis-coded quantization decoder (the QTIP / EXL3 quality tier) to run outside CUDA. See the repo for the quantizer, the WGSL kernels, and the full verification harness. ## Quality (WikiText-2 perplexity, no fine-tuning) | | fp16 | **TCQ 3-bit (this model)** | vs fp16 | |---|---|---|---| | Qwen3-8B | 9.02 | **10.34** | 1.15× | Trellis-coded quantization is the current quality frontier for low-bit LLM weights, beating scalar quantization (GPTQ / AWQ / GGUF) at equal bitrate. The K=2 quantizer in this project reproduces the QTIP paper's published rate–distortion (MSE 0.0739 vs 0.0733). ## This packed model - **36 layers**, packed to **~3 bits/weight** (2.9 GB on disk). - runs in a browser on a **4 GB GPU** — verified live in Chrome on a laptop **RTX 3050 Ti (4 GB)**, generating the completion above without running out of memory (256-token KV cache; `maxSeq` is configurable for longer contexts). - Verified: the full model runs through the exact shipping WGSL shaders and generates coherent, factually-correct text; kernels match NumPy to `<1e-6`; 135M logits are top-5 exact vs PyTorch. ## Format This is **not** a standard `transformers` checkpoint. It is a packed 3-bit format (`manifest.json` + sharded `.bin` weights + IP metadata) designed for the WebGPU runtime in the [Trellis WebGPU](https://github.com/Azimml/trellis-webgpu) repo. To run it: ```bash git clone https://github.com/Azimml/trellis-webgpu # place these files under web/model_8b/ , then: cd web && python3 -m http.server 8000 # open index.html in a WebGPU browser # or verify headlessly on your GPU: python scripts/run_packed_headless.py web/model_8b "The capital of France is" 40 ``` ## License & credit Quantized derivative of [Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B) (Apache-2.0), distributed under the base model's license. Method: QTIP (Tseng et al., NeurIPS 2024) and EXL3 (turboderp). Quantization + WebGPU port: independent reimplementation.