Qwenputer renders a single frame of Doom using a CPU inside LLM weights

Developer divagr18 created Qwenputer, an experimental build of Alibaba’s Qwen3.5-4B that has a 32-bit RISC-V CPU compiled directly into its weights. The CPU uses 11 transformer layers placed in front of the model's stack. Instead of using statistical guesses, a compiler calculates every value so the layers behave as logic. The project uses the model's memory state as RAM by utilizing Gated DeltaNet layers to store and read data.

Doom · official art via IGDB

Developer Aditya Narayana recently used the setup to render a single 320x240 frame of Doom. The process used 6,342,933 trace tokens and took 50.7 minutes on an RTX 5070 Ti 16 GB, though a fused executor later reduced that time to about 24 minutes. This run omitted pretrained Qwen language layers and relied on a specialized Triton execution tool rather than standard Hugging Face software.

Divagr18 says the system can run C and MicroPython and has passed official riscv-tests suites. The CPU integration helps with arithmetic; Qwenputer scored 97.5 on DeepMind’s big-number multiplication set, while the base Qwen3.5-4B scored 55 out of 100. Reports say the Doom run produced only one frame rather than actual gameplay, and interactive frame rates are not expected.