Developers embed a 32-bit CPU into Qwen’s model weights, enabling the large language model to run computer programs directly internally.

Beating AI Express News: Independent developer Divyansh Agrawal has built Qwenputer based on Alibaba’s Qwen3.5-4B, integrating a 32-bit RISC-V virtual CPU directly into the model’s weights. This allows the large language model (LLM) to perform precise calculations when generating responses without relying on external computing tools. Instead of retraining Qwen, Agrawal added 11 layers before the original 32-layer network, using a compiler to translate the CPU’s operation rules directly into model parameters. The added network leverages the Gated DeltaNet architecture to implement precisely readable and writable memory for storing programs and data. The virtual CPU can now run compiled C programs and has passed RISC-V instruction tests, with no errors detected in tests executing approximately 18 million total instructions. In DeepMind’s 40 large-number multiplication problems, the original Qwen got 22 correct answers, while Qwenputer scored 39; the original Qwen achieved a perfect score when connected to external Python tools. The CPU itself executes instructions accurately, but the LLM may still make errors when calling it or outputting results. This technical approach has precedents: in March this year, Percepta implemented a virtual computer capable of running C programs using a neural network, though it was a purpose-built model. By contrast, Qwenputer integrates similar computing capabilities into the already trained Qwen, enabling the LLM to use its internal CPU for calculations during response generation. However, the solution is currently very inefficient: a large number of zero-value parameters expand the model file to 18.8GB, roughly twice the size of the original, and its computing speed is far slower than external tools.