div (fp32)

1/4

Meet Qwenputer: Qwen3.5-4B with a full RISC-V CPU inside its weights, using the model's own Gated DeltaNet layers as RAM.

It runs C and MicroPython in its forward passes, no external runtime.

On DeepMind math: base 55–90%, base + Python tool 95–100%, Qwenputer 97.5–100%.

Told my agent to eliminate all implicit state machines and it put a hit out on me

Added more GPU compute, try some RL runs out live!

Made Nemesis for @AIatMeta: a real-to-sim robotics RL platform. Turn real places into simulations, break robot dogs and humanoids then retrain them with PPO, let Muse agents hunt for failures on Autopilot, compete against Muse on breaking robots faster.

nemesis.divagr.com

Some fun stuff I'm considering for Qwenputer

1) Getting multiple GPUs to run inside one model, with like scatter/reduce and parallel execution, maybe some form of tensor fusion
2) Getting it working on IoT/mobile devices for edge inference computation without a bundled external runtime
3) Really fun possibility - moving CPU-LLM communication into latent space (DMs open if you've worked on something similar and have concrete ideas!)

as soon as I CAN GET MY HANDS ON A DAMN 5090