Artificial intelligence / news brief
qwenGPU compiles a small SIMT processor into Qwen3.5-2B weights
- Event
- Published

An independent developer released qwenGPU, an MIT-licensed experiment that inserts 45 compiled transformer layers ahead of Qwen3.5-2B to execute a limited CUDA subset through model state. The repository reports exact replay for its launches and agreement with a Tesla T4 on 49 test kernels, but the tests are creator-run and the model produces only 10–15 tokens per second on Apple MPS.
Why it matters
The project explores deterministic computation inside model weights, but it is a research demonstration rather than a practical GPU replacement or official Qwen release.
Sources
- qwenGPU repository (external site)Vixhal Baraiya · primary source ·
- Meet qwenGPU: Qwen3.5-2B With 45 Compiled GPU Layers That Run CUDA (external site)MarkTechPost · coverage ·