Qwenputer compiles an exact 32-bit RISC-V CPU into Qwen3.5-4B weights, with 83.3% multi-session recall
solyarisoftware · x · 2026-10-29
Qwenputer is an experimental open-source model that compiles an exact 32-bit RISC-V CPU directly into Qwen3.5-4B's weights using 11 extra transformer layers — none learned; every weight is computed by a compiler. The model's Gated DeltaNet state acts as RAM, letting the model run calculations internally instead of calling Python or external tools.
An experimental version adds a filesystem: up to 512 files (53K tokens) stored in DeltaNet heads with put/append/get/list/delete operations and 2-3 token directory lookups. Save and reload the state and files persist across chats. The developer reports 83.3% accuracy on a 90-question multi-session recall test vs 7.8% for base Qwen3.5-4B.
Code and weights are on GitHub (141 stars) and Hugging Face.
Related event: Qwenputer compiles a 32-bit RISC-V CPU into Qwen3.5-4B weights(7 posts)→
More from Research
- AI Now Discovers Faster Than Institutions Can Verify: The 200M-Protein Integrity Paradox — mikeflache · 2026-10-29
- FrontierCS 2 reshuffles rankings: Claude Opus 5.5 tops FCI as DeepSeek 1017 degrades — teortaxesTex · 2026-10-29
- Constitutional grafting: character training for open models under $500, skipping $440k of compute — voooooogel · 2026-10-29
- Transformer layers turned into a full-fledged symbolic CPU — yshan2u · 2026-10-29
- NanoDreamer: A 750K-parameter world-model agent trained on a MacBook M5 Pro plays 5 MinAtar games — CatAstro_Piyush · 2026-10-29
- Why You Can't Just Remove Positional Embeddings: Self-Attention Is Permutation Equivariant — cneuralnetwork · 2026-10-29