Local Qwen 3.8 27B reads raw model weights
Original: Qwen 3.8 27B on my laptop can interpret model weights, too
Why This Matters
Weight-space reasoning in local models could change how ML debugging is done.
Simon Willison ran the weight-interpretation experiment from a Nov 2026 paper on his laptop using Qwen 3.8 27B, reproducing results previously attributed only to frontier models like GPT-6 Astra—no tools, no code execution.
A paper titled 'Sparks of Weight-Space Programming' by Takehi, Teoh, Zhang, and Khattab found that GPT-6 Astra can read and modify raw neural network weights—floating-point numbers representing a trained model—and deduce what that model was trained to do. The authors call this 'Weight-Space Programming.' Key claims: Astra identified a reasoning-loop bug in Qwen3.5-0.8B's weights and fixed it by editing just 0.016% of them, lifting MATH-500 scores from 7.9% to 48.1% and tripling IFEval performance—all without writing a training loop.
Willison tested the simpler 'which game?' experiment from the paper's GitHub repo, where an LLM is handed raw weights of a network trained to play one of four Atari-style games and asked to identify it with zero tool access. His local Qwen 3.8 27B pulled it off. The finding matters because it suggests weight-interpretation isn't exclusive to top-tier API models.
Not everyone is convinced the paper's framing holds up. spaCy author Matthew Honnibal called it misleading for conflating no-tool weight reading with tool-assisted model editing. Sean Goedecke argued the Qwen fix example just replicates steps a human ML researcher would take anyway. Willison himself separates the two claims, focusing on the no-tools case as the genuinely novel question.