divagr1925 posted on X about Qwenputer, a build that puts a full RISC-V CPU inside the weights of Qwen3.5-4B. The model’s own Gated DeltaNet layers serve as RAM, so C and MicroPython run during the forward passes with no external runtime.
Qwen3.5 uses Gated DeltaNet in three of every four layers. divagr1925 turned those layers into memory using one-hot keys to pick addresses. With the gate and write strength set to 1, individual slots can be overwritten and read back exactly, and the state stays a fixed size as the chat grows.
On DeepMind math, base Qwen scored 55 to 90 percent, base plus a Python tool scored 95 to 100 percent, and Qwenputer scored 97.5 to 100 percent. On NUPA, Qwenputer hit 100 percent against 48.6 percent for the base model.
There is also a filesystem living in the GDN state. Save the state after a chat, load it into the next one, and the files are still there.…