Qwenputer Puts a Working CPU Inside a Large Language Model

Regular AI models predict the next word to answer questions, so they often mess up big math like multiplying large numbers, and usually need an outside calculator or code tool to help.

A developer named Divyansh Agrawal released an experimental model called Qwenputer, built on Alibaba's Qwen3.5-4B. It packs the RISC-V instruction set into the model weights and uses the model's internal states as RAM, creating a working CPU inside the neural network.

In tests, Qwenputer pushed accuracy on large number multiplication from 55% to 97.5%, and scored 100% on the NUPA long math benchmark, while the standard model only got 48.6%. It has run about 18 million instructions without errors.

The catch: the model grew to about 18.8GB, nearly double the original, and a single CPU step can take seconds. The project is open source on Hugging Face and GitHub, and runs on one GPU with 24GB of memory.