Qwen 3.8 27B on my laptop can interpret model weights, too

What happened: A blog post reports that Qwen 3.8 27B, run locally on a laptop, identified which game four small policy networks were trained to play by reading their raw weights, following a paper on Weight-Space Programming. Key numbers: Qwen 3.8 27B is 17GB. Why it matters: The post focuses on the no-tools question of whether models can understand behavior from weights. spaCy author Matthew Honnibal called the paper misleading for conflating that with code-assisted model updates, and Sean Goedecke deconstructed the Qwen3.5-0.8B example.

A blog post reports that Qwen 3.8 27B, run locally on a laptop, identified which game four small policy networks were trained to play by reading their raw weights, following a paper on Weight-Space Programming.

The post focuses on the no-tools question of whether models can understand behavior from weights. spaCy author Matthew Honnibal called the paper misleading for conflating that with code-assisted model updates, and Sean Goedecke deconstructed the Qwen3.5-0.8B example.