MiniMax M2.7

This is the strongest sustained decode case, with 83.2 tok/s median throughput over 4096 average output tokens. The long output did not materially degrade throughput, indicating good steady-state generation performance.

Performance here is mid-pack at 68.2 tok/s median throughput with similarly long outputs around 3953 tokens. The result is stable, but notably slower than the API Design workload, suggesting sensitivity to prompt/content structure.

This is the weakest throughput case at 40.2 tok/s median, despite a shorter average output of 2907 tokens. The drop implies the model slows materially on more open-ended, prose-heavy generation.

TTFT is slow at 6325 ms median, and the low average output token count of 277 means startup latency is not being amortized by long generations. This is a poor interactive latency profile for short-answer tasks.

TTFT improves somewhat to 5384 ms median, but remains high for a factual prompt with only 163 average output tokens. Startup latency is still the dominant bottleneck rather than decode speed.