Meet qwenGPU: Qwen3.5-2B With 45 Compiled GPU Layers That Run CUDA

| Source: MarkTechPost

Tags: qwenGPU, Qwen3.5-2B, CUDA, SIMT, Gated DeltaNet

qwenGPU compiles a 4-lane SIMT GPU into Qwen3.5-2B as 45 untrained layers, letting the model execute real CUDA launches step by step inside its own weights.

Details

qwenGPU adds 45 hand-compiled, untrained layers to Qwen3.5-2B that implement a small 4-lane SIMT GPU. The model writes out every step of real CUDA launches itself, using Gated DeltaNet state as registers and memory. It is an experiment in compiling programs directly into transformer weights rather than a practical accelerator.