Qwenputer embeds a fully functional RISC-V CPU inside Qwen3.5's Gated DeltaNet attention layers, using them as RAM. It can execute arithmetic operations, C code, and MicroPython entirely within the model's forward passes without external tools or runtime.
Background
Qwen3.5 uses Gated DeltaNet, a linear RNN, as its primary attention mechanism in 3 out of 4 layers, where each head maintains a state matrix. This research explores repurposing those state matrices as programmable memory for an embedded CPU.
- Source
- Lobsters
- Published
- Oct 29, 2026 at 06:00 PM
- Score
- 8.0 / 10