You can get small models to be really good at math benchmarks, but they are no longer language models in this case. They become math models. We had a point around a year ago where the models where very general and now we go back to specialized models.

https://pretty-radio-b75.notion.site/DeepScaleR-Surpassing-O1-Preview-with-a-1-5B-Model-by-Scaling-RL-19681902c1468005bed8ca303013a4e2