Qwen 3.8 27B can interpret and modify neural network weights

Simon Willison reported testing a local capability echoing a finding first described by Rikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab in a paper-shaped blog post titled “Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights,” published the previous day. That work found frontier models, specifically GPT-6 Astra, capable of both interpreting and modifying the raw weights of a machine learning model, using a text description to operate directly on numerical parameters rather than treating them as opaque binary data.

Willison’s post, published November 3, 2026, under the title “Qwen 3.8 27B on my laptop can interpret model weights, too,” extends the original finding by demonstrating the behavior on a smaller, locally run model. Willison’s title indicates that Qwen 3.8 27B, run on his own laptop rather than on frontier infrastructure, exhibited comparable weight-interpretation ability.

The original authors’ framing, as paraphrased by Willison, centers the significance of the finding on models treating their own or another model’s weights as readable, writable content rather than as fixed numerical artifacts. Willison’s follow-up suggests this is not confined to the largest, most resource-intensive systems but is reproducible on consumer hardware with a smaller model.

No further technical detail, methodology, or quoted text from either the original paper or Willison’s post is given in the available material beyond these summary points.