Qwenputer combines the stock Qwen 3.5 4B large language model with a 32-bit RISC-V CPU whose implementation is embedded directly into the model’s weights. The experimental system can perform processor arithmetic and execute C and MicroPython code without relying on an external runtime or tool call. It can also run DOOM, although at very low speed.
The CPU is implemented through 11 transformer layers placed before Qwen’s own layers. Rather than being trained, their weights are calculated by a compiler. According to Agrawal, the processor passes the official RISC-V tests, produces programs matching QEMU, and has executed 18 million instructions without an error.
Qwenputer uses Qwen’s Gated DeltaNet state as RAM, allowing the model to store and retrieve instructions internally. Agrawal described the use of these “fast weights” as RAM as an established idea, while presenting its application to Qwen 3.5 as a first for this project. The built-in processor also enables exact calculations, addressing a common weakness of language models.
The project includes a filesystem that also relies on the DeltaNet state. Its storage provides 1,024 slots for up to 512 files, and the contents can persist when the model state is saved at the end of a conversation.
These additions have nearly doubled the size of the underlying model, bringing Qwenputer to almost 19GB. Agrawal recommends a workstation with at least 24GB of VRAM for testing the current version. Work is continuing to improve its coding workflow and performance.
Qwenputer remains an experimental project focused on exploring the possibilities of integrating computing functions into an AI model. Its code is available under the MIT license on GitHub, and the model can be downloaded from Hugging Face.
Sources: Tom's Hardware.