qwenGPU compiles a GPU into Qwen3.5-2B's weights to run CUDA in forward passes
AIDeveloper Vixhal released qwenGPU, an experimental Qwen3.5-2B model with 45 compiled GPU layers placed before its 24 original layers, under the MIT license on GitHub. The model runs CUDA compiled to NVIDIA PTX inside its forward passes, and the developer reports every launch was exact with a text drift KL of about 7e-4.
Source details
MarkTechPost · Michal Sutter · marktechpost.comPublished · added here
Original title: Meet qwenGPU: Qwen3.5-2B With 45 Compiled GPU Layers That Run CUDA