That was a long grueling test ... Threw the same question at qwen 3.8:28 fully in vram, it smoked it answer wise... And took 10 seconds versus an hour for 500 tokens 😂 but it was a good experiment and I got a good blog write up on it .