MiniMax-M2.7 is highly verbose and slow

[sweihub]

I upgraded from M2.5 to M2.7, I am using 4 x A100(80G), the M2.5 is compared to be on-the-fly, but the M2.7 is really verbose and slow, I am running with the VLLM, it's getting worse after some amount of requests.

[sweihub]

I have to use --quantization fp8 to speed it up, thanks anyway for the awesome model!

[sweihub]

Due to the verbose thinking issue, the Minimax M2.7 kills my Claude code with insane context, the context grows rapidly, and the claude code is super slow, any suggestion?

[sweihub]

No official response from the staff?
I've figure out a system prompt/CLAUDE.md to make it thinking concisely, hopefully!

System Prompt

Constraint: Limit internal logic to <100 words.
Process: Audit as a senior expert; prioritize first-principles and logical consistency.
Output: Professional, high-density results only. Maximize word-to-value ratio by stripping filler, hedging, and redundancy.

[sweihub]

Blood and tears, the fp8 quantization will make the minimax m2.7 a dumb.
The bug created by minimax m2.7 bf16 was fixed by Qwen3.6-35B-A3B-FP8.