Qwenputer fuses a RISC-V CPU into Qwen 3.5 weights and can run Doom
The model executes C and MicroPython internally without external tools, but its enlarged size of nearly 19GB sets a high hardware bar.
An AI engineer, Divyansh Agrawal, has released Qwenputer, an experimental build that fuses the 4-billion-parameter Qwen 3.5 model with a 32-bit RISC-V processor whose weights were produced by a compiler rather than by training. The processor sits as eleven layers ahead of Qwen's own, and the model can execute C and MicroPython code internally without calling any external tool or runtime, giving exact arithmetic where language models usually struggle. Agrawal repurposed the model's Gated DeltaNet state as working memory and, on that same foundation, added a 1,024-slot filesystem holding up to 512 files, which keeps its contents only when the state is saved at the close of a conversation. The enlarged model is close to 19GB, so Agrawal suggests a workstation with at least 24GB of VRAM, and the project, released under the MIT licence, can run Doom though slowly.
Why it matters
For readers following AI, the project shows that exact computation can be embedded directly in a language model's weights rather than delegated to an outside tool, a different route from the usual pattern of models calling separate runtimes. The results are still experimental and slow, and the file store survives only when the model state is saved at the end of a conversation, so the approach is better understood as a research demonstration than a settled method.
Matters or hype?
Does this story matter, or is it hype? Vote before you see what everyone else said.