QwenPU runs Karpathy's stories260K on an NPU compiled inside Qwen 3.5-4B's weights
LatentSpacer · reddit · 2026-11-01
divagr18, who previously put a RISC-V CPU inside Qwen, now presents QwenPU: Karpathy's stories260K running on an NPU compiled into Qwen 3.5-4B's weights—literally an LLM inside an LLM.
- stories260K is just a C program; the author compiled an NPU into Qwen's weights that performs matmuls, softmax, and attention bit-exactly, with the stories260K weights living in Qwen's own memory and no external tools
- The screenshot shows the vector unit inside Qwen's DeltaNet state (VMAT/VEW/VQ instructions, 64 int8 lanes) and step-by-step traces comparing Plain C int8, Qwenputer NPU, and QwenPU
- Code and weights are on GitHub and Hugging Face; the author is ill, so the full writeup, optimizations, and size reductions are still in progress
More from Research
- After42: experts rate 45 of AI's 135 proposed research questions genuinely interesting — james_y_zou · 2026-11-01
- Google tests two screening policies to cut noise from agentic automated bug repair — burkov · 2026-11-01
- AI novelty checks reject 96% of ideas human scientists later publish, UIUC paper finds — rohanpaul_ai · 2026-11-01
- Why the pinky's position matters for robot dexterity: thumb as fulcrum — chris_j_paxton · 2026-11-01
- Princeton's Mengdi Wang to unveil LabOS, an AI-XR co-scientist that works in real labs — lihua_lei_stat · 2026-11-01
- ScienceClaw AI-for-science hackathon kicks off at MIT Media Lab — ProfBuehlerMIT · 2026-11-01