GoKawiilDeveloper embeds RISC-V CPU into Qwen 3.5 4B LLM's weights, runs DOOM on it
AI engineer Divyansh Agrawal built 'Qwenputer,' a modified version of Qwen 3.5 4B that has a 32-bit RISC-V CPU compiled directly into 11 additional transformer layers placed ahead of the model's own weights. The CPU uses the model's Gated DeltaNet state as RAM and can execute C and MicroPython code and arithmetic operations without calling any external tools, and has been shown running the game Doom, albeit slowly.
GoKawiil's interpretation of the reporting above, not reported fact.
Embedding a verified CPU inside an LLM's own computation could address a long-standing weakness of language models: unreliable arithmetic and code execution, since calculations here are handled by compiled logic rather than learned weights. Agrawal's claim that the CPU passes official RISC-V test suites and matches emulator behavior suggests a path toward LLMs that perform exact computation internally rather than relying on external interpreters or tool calls, though this remains an experimental proof of concept rather than a production technique.
- Qwenputer embeds a compiler-generated, non-trained RISC-V CPU into 11 extra transformer layers ahead of Qwen 3.5 4B's own weights.
- The embedded CPU uses the model's Gated DeltaNet state as memory and has run Doom's E1M1 level, plus C and MicroPython code, without external tool calls.
- Agrawal says the CPU passes official riscv-tests and matched QEMU output over 18 million error-free instructions, and plans to open-source the project.
Source: tomshardware.com — Mark Tyson, 2026-11-02
Published there as: “AI model gets a 32-bit RISC-V CPU compiled directly into its weights, then runs DOOM on it — experimental AI model can run Doom and C code using its own weights as RAM”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.