1.5k points280 comments
Astra edits LLM weights directly, boosting Qwen math scores
What is Astra can read and write weights to improve Qwen?
A new model called Astra can examine the raw numbers inside another neural network and modify them, without any training loop. By tweaking just 0.016% of Qwen-3.5-0.8B’s weights it lifts its math benchmark from 7.9% to 48.1% and triples its instruction-following score.
Who it is for
Researchers and developers focused on model optimization or interpretability can use Astra to patch or fine-tune models without costly gradient computation.
How to try it
Visit the blog’s GitHub link for code and demo scripts