According to CoinMeta, independent developer Divyansh Agrawal developed Qwenputer based on Alibaba’s Qwen (3.5-4B), incorporating a 32-bit RISC-V virtual CPU into the model weights, allowing large models to perform precise calculations without the need to call external tools. This virtual CPU can run compiled C programs and has been tested through RISC-V instructions; approximately 18 million instructions were executed without any errors being found. In 40 large-number multiplication problems in DeepMind, Qwenputer answered 39 correctly, while the original version Qwen only answered 22 correctly. Although CPU can execute instructions accurately, large models may still make mistakes when calling or outputting results. There are precedents for this technical approach; in March of this year, Percepta implemented a virtual computer that could run C programs using neural networks. Qwenputer integrates similar computational capabilities into already trained Qwen, allowing language models to use internal CPU when generating answers. However, this solution is less efficient, with the model file size reaching 18.8GB, which is about twice that of the original version, and its computing speed is also not as fast as that of external tools.