Qwenputer runs a full RISC-V CPU inside Qwen3.5 weights

Researcher Divyansh Agrawal built Qwenputer, a 4-billion-parameter Qwen3.5 model that executes a complete RISC-V CPU with no plugins or external sandbox. It exploits Gated DeltaNet, the linear attention layer Alibaba uses in Qwen3-Next and Qwen3.5: with the gate fixed at 1.0 and one-hot keys, the attention state becomes exact, addressable memory.

• Qwenputer compiles an RV32IM CPU into 11 of Qwen3.5's 32 layers without extra training
• Per-head attention state stays fixed-size regardless of context length, unlike KV-cache
• By default the gate is below 1.0, so memory fades instead of being stored permanently
• Gated DeltaNet occupies 3 of every 4 attention layers in Qwen3-Next