3 ms·

This is obviously a continuation training of 3.5, it's not a new model architecture but an incremental improvement.