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

Researchers discovered that large language models, including smaller models like Qwen 3.8 27B running locally, can interpret and analyze raw neural network weights to understand what a model was trained to do—a capability termed "Weight-Space Programming." The paper demonstrates that frontier models like GPT-6 Astra can not only read model weights but also modify them to improve performance, such as fixing a Qwen model's math reasoning by editing just 0.016% of its weights and raising its MATH-500 score from 7.9% to 48.1%.