Qwenputer is a stock Qwen 3.5 4B model with a RISC-V CPU embedded directly into its weights. The CPU uses the model's own Gated Delta Net (GDN) attention layers as RAM, running arithmetic operations, C, and Micro Python code entirely within forward passes—no plugins, tools, or sandboxes required. By setting GDN gates to 1.0 and write strength (beta) to 1.2, the GDN state becomes exact and addressable memory.
One-hot key encoding allows precise row reads and writes, enabling the GDN layers to function as RAM at a lower cost than traditional attention KV caches. No training is needed; a compiler sets every weight directly. The project also includes experimental in-model file storage and two follow-up papers on hidden reasoning in RNN states and tamper-proof monitoring tools.
Qwenputer is available on Hugging Face and Model Scope, with code on GitHub.
Source: Hacker News · Summarized by HeadlinesBriefing