More parameters = more facts stored. Knowledges are almost incompressible, where strong reasoning only requires a 3B core or so.
More parameters = more facts stored. Knowledges are almost incompressible, where strong reasoning only requires a 3B core or so.
Weibo's VibeThinker manages with half of that: https://arxiv.org/abs/2511.06221 (They finetuned Qwen2.5-Math-1.5B for reasoning.)