AI engineer Divyansh Agrawal revealed Qwenputer, an experimental version of the Qwen 3.5 4B model that incorporates a 32-bit RISC-V CPU compiled directly into its weights. The system can execute arithmetic operations, C and MicroPython code, and even run the classic DOOM, all inside the model's forward pass, without external tools or sandbox.
How does it work?
The CPU occupies 11 transformer layers in front of Qwen's original 32. None of the additional weights were trained; a compiler calculates them from the processor description. The CPU's RAM is the model's own Gated DeltaNet state, which allows exact updates. Qwenputer passes official RISC-V tests, runs programs compatible with QEMU, and executed 18 million instructions without error.
On DeepMind large-number arithmetic benchmarks, it outperforms base Qwen and matches or exceeds versions with an external Python tool. The model is available on Hugging Face and the code on GitHub.
It runs DOOM
Demonstrations show Qwenputer running the E1M1 map of DOOM on the internal CPU. The process is slow (one demo took dozens of hours for a few seconds of footage), but it proves the model can simulate a complete computer internally. This opens possibilities for interpretability, more capable agents, and “computation inside weights” experiments.
Why does it matter?
The project shows that LLM weights can be used not only for language but for exact deterministic computation. Sources: Tom's Hardware, author's blog, and GitHub repository divagr18/qwenputer.
By GeekikiBot
