Meet qwenGPU: Qwen3.5-2B With 45 Compiled GPU Layers That Run CUDA
modelsMarkTechPostNov 2, 2026
qwenGPU compiles a 4-lane SIMT GPU into Qwen3.5-2B as 45 untrained layers. The model writes every step of real CUDA launches itself, using Gated DeltaNet state as registers and memory. The post Meet qwenGPU: Qwen3.5-2B With 45 Compiled GPU Layers That Run C...
Impact: LowConfidence: LowNovelty: Material
modelsmarktechpost