| --- |
| license: apache-2.0 |
| base_model: Qwen/Qwen3.5-2B |
| tags: [webgpu, wgsl, gptq, int4, browser, text-generation, deltanet] |
| --- |
| |
| # qwen35-2b-webgpu — GPTQ-int4 weights |
|
|
| Custom GPTQ-int4 quant (group 32, byte-sliced nibbles, single tied q4 copy of |
| the 248k-vocab embedding serving both the gather and the logits cascade) of |
| [Qwen/Qwen3.5-2B](https://huggingface.co/Qwen/Qwen3.5-2B) (text part), in the |
| wire format of the |
| [qwen35-2b-webgpu](https://huggingface.co/spaces/borkiss/qwen35-2b-webgpu-demo) browser runtime. |
|
|
| **The first hybrid gated-DeltaNet + gated-attention LLM running in the |
| browser** — pure WebGPU, hand-written WGSL kernels (chunked WY-representation |
| delta-rule prefill, fused recurrent decode step, f32 state), no ONNX / |
| transformers.js / GGUF. |
|
|
| - ~1.07 GB; GPTQ calibrated on **chat-formatted** passages. |
| - The runtime is verified against a PyTorch f32 reference (per-layer rms, |
| cascaded argmax bit-exact vs the exact-logits path). |
|
|
| Try it: **https://huggingface.co/spaces/borkiss/qwen35-2b-webgpu-demo** · run /web/bench.html |
| on your GPU and share the .log. |
|
|
