This is clearly what is happening. Deepseek can train on o1 generated synthetic data and generate a very capable and small model. This requires that somebody build an o1 and make it available via API first.