Qwenputer compiles a small CPU into model layers
Qwenputer adds a compiled RISC-V processor to Qwen3.5-4B, using the model’s recurrent attention state as addressable memory. Arithmetic executes through added weights and decoding steps rather than an external calculator service. Divyansh Agrawal reports passing processor tests and running C and MicroPython, but the experiment needs a custom driver and performs calculations far more slowly than an ordinary CPU.
Divyansh Agrawal · Oct 22, 2026
Why it matters
The project offers a concrete experiment in combining exact computation with language generation inside one model execution path.
More context
The original article is dated October 22. Code and checkpoints are public; the documented setup uses uv, the Hugging Face CLI and one 24 GB GPU. Start by inspecting the processor tests and driver, then reproduce a small arithmetic example.
Keep in mind
The 18.8 GB checkpoint and custom RvChat driver do not work through a plain generate call. Processor and arithmetic results are author-reported; this is an experiment, not a general replacement for external tools.
Original read