2 ms·

better prompt processing like 1.5x+ and more kv but tg most likely lower like 0.8x or so but I am just going by memory for Qwen3.5 without mtp.