Developer compiles a vector unit into Qwen3.5-4B's weights to run a tiny Llama inside the model

QwenPU adds a vector unit to the Qwenputer project's compiled RISC-V CPU so Karpathy's 260K-parameter stories260K runs as a C program during forward passes; no new weights are trained, and the work-in-progress 26.7 GB checkpoint needs a 40 GB or larger GPU to load whole.