qwenTPU compiles a TPU into Qwen3.5 weights, achieving exact integer matmul
CatAstro_Piyush · x · 2026-11-04
Developer Bobliuuu open-sourced qwenTPU: a weight-stationary matrix-multiply array (up to 32×32, 1,024 MACs per matvec, 5-bit operands) compiled into stock Qwen3.5 (0.8B/2B/4B) via new layers, with no TPU hardware and no training — Qwen's own weights untouched.
Key results:
- Integer-exact output vs NumPy: 0 mismatches across 1,139,712 replayed token positions on four machines
- Plain text generation unaffected (KL divergence 3e-4 to 8e-4)
- Two versions: v1 does exact matmul; v2 runs a real neural network program through the same weights, using Gated DeltaNet layers as memory and registers
The project continues the trend of embedding CPUs, GPUs, and filesystems inside LLM weights. MIT-licensed on GitHub with a demo video and write-up.
More from Research
- Anthropic says it may imminently hit RSP threshold for AI-accelerated R&D, releases provable inference work — tobyordoxford · 2026-11-04
- Google scanned 1,033 real household products, packed into a 345 MB GLB for three.js — victormustar · 2026-11-04
- These 10 GPU optimization papers are why your model runs fast today — NandoDF · 2026-11-04
- INT8 isn't the problem: I benchmarked samplers on INT8 flow models — YakovSava · 2026-11-04
- Karpathy Boosts Open Hacker News Dataset: 10M+ Posts Usable as a Personal Knowledge Base — huangyun_122 · 2026-11-04
- PixelCraft: A Multi-Agent Framework Targets Multimodal AI's Chart-Reading Failures — burkov · 2026-11-04