2 ms·

Probably because they are loading the entire model into SRAM. Thats how they can achieve 1.5k tokens/s.