Qwenputer prototype embeds a processor in an AI model to run code

AI engineer Divyansh Agrawal has released Qwenputer, an experimental modification of Qwen 3.5 4B that runs C and MicroPython code through a processor embedded in the model, without external runtime calls. The model is available on Hugging Face, with MIT-licensed code on GitHub. A compiler generates 11 additional transformer layers implementing a 32-bit RISC-V processor. The model uses its DeltaNet state as working memory and filesystem storage. Agrawal says the processor passed official RISC-V tests and ran 18 million instructions without an error. The modified model occupies almost 19GB; Agrawal recommends at least 24GB of video memory for testing.