Qwenputer integrates risc-v CPU inside model for enhanced arithmetic capabilities

On October 30, 2026, IT Home reported that developer Divyansh Agrawal released an experimental model called Qwenputer on October 22. The model, built on Alibaba's Qwen3.5-4B, embeds a 32-bit RISC-V CPU and memory within its neural network, allowing large language models to perform calculations without external tools. Traditional AI models often struggle with complex arithmetic tasks, requiring external calculators or code interpreters. Qwenputer, however, compiles the RV32IM instruction set into model weights and uses Qwen3.5's gated DeltaNet state as RAM to simulate a CPU. Testing showed significant improvements: accuracy in large integer multiplication rose from 55% to 97.5%, and division from 72.5% to 97.5%. In the NUPA benchmark, Qwenputer achieved 100% accuracy, compared to 48.6% for the standard version. The model also includes an experimental in-model file system, storing up to 512 files and 53,000 tokens, with a 90% accuracy rate in recalling prior information. However, the model's size increased to 18.8GB, and execution is slower due to the need to generate multiple tokens per instruction. The project is now open-sourced on Hugging Face and GitHub.