Weight-space programming: agents edit LLM weights instead of training
Editing 1,023 weights lifts a Qwen model 29 points on MATH-500, no training; plus DeepSeek's self-hostable agent runtime reportedly passes 1.5M daily users.

Copy markdown
Tuning by editing weights, not training
A new Hugging Face paper has a reasoning agent inspect and directly rewrite a model's parameters instead of running gradient descent. Editing just 1,023 weights of Qwen3.5-0.8B lifted MATH-500 by 29.45 points, with gains on IFEval (+36.72) and SciBench (+17.11) and some transfer to unseen tasks. It's early and uneven — 6 of 16 runs couldn't validate an improvement and kept the original weights, and no code is out yet — but it's a concrete look at adapting models by surgical edits rather than retraining.
Update: DeepSeek's open agent runtime hits frontier scale
DeepSeek's Harness — the open, self-hostable agent app — has reportedly crossed 1.54M daily active users and ~92T tokens/day less than three months after launch, per Bloomberg. For builders the signal isn't the headline number: it's that an agent runtime you can run yourself is now proven at that volume, de-risking it as a base to build your own agents on.