A developer compiled a working computer into a Qwen model's weights

Divyansh Agrawal built Qwenputer: Qwen 3.5 4B with a 32-bit RISC-V processor written directly into 11 added layers, no training involved, using the model's own memory state as RAM. It runs C and MicroPython inside an ordinary forward pass. On long-integer arithmetic it scores 100% against the base model's 48.6%, and ordinary text output stays almost identical to Qwen's.

It is slow, about 3,200 processor steps a second on an L4 GPU, and the checkpoint is 18.8 GB. As a product it is a curiosity. As a proof it shows that a language model's layers can be made to compute exactly, and exactness is what language models have never had at arithmetic. SourcesAA

Sources

• AQwenputer (Divyansh Agrawal)divagr.com
• Adivagr18/qwenputer (GitHub)github.com

This story ran in the Morning Brief of October 29th 2026.