MiniMax M2.5 hits about 80% on Sweetbench and runs near 100 tokens per second, helping teams deploy faster models on tighter budgets.