Qwenputer compiles a full RISC-V CPU into Qwen3.5-4B weights, bit-exact, even running another LLM
udmrzn · x · 2026-11-03
A developer compiled a complete 32-bit RISC-V CPU directly into Qwen3.5-4B weights — dubbed Qwenputer — and even ran another LLM (Karpathy's stories260K) inside it, all within the model's own forward pass with zero external runtime or tool calls. Key details:
- Compiled CPU: an 11-layer RV32IM (with Zba and bignum unit) sits in front of Qwen; weights are compiler-generated, no training.
- Model as RAM: Gated DeltaNet state serves as exact read/write memory holding registers, PC and stack.
- CPU → NPU: QwenPU adds compiled vector units (matmul, softmax, attention), running stories260K as a C program inside the model.
- Results: DeepMind bignum multiplication benchmark jumps from 55% to 97.5%+, partly beating the base model with Python tools; fully bit-exact over 230k+ steps.
Fully open source (MIT code, Apache-2.0 weights) with quickstart commands included.
More from Research
- Recursive GPT-5-Mini calls often beat a single GPT-5 call for less money — lateinteraction · 2026-11-03
- Same Agent Reimplements Liquid in Rust vs C: Rust Uses 42% Fewer Tokens, C Runs 1.7x Faster — antirez · 2026-11-03
- After viral 'fly plays Beat Saber', researchers call for shared whole-brain emulation benchmark — rishabh16_ · 2026-11-03
- Researchers uncover 120 Google search twiddlers, including AI Mode's fan-out plan — VeryWellVersed · 2026-11-03
- Astra IDs Breakout, Pong, Pac-Man and Freeway from raw model weights alone — burny_tech · 2026-11-03
- World model that predicts the future directly in 3D, fully open-sourced — YunzhuLiYZ · 2026-11-03