4 × NVIDIA H100 (94 GB), one node of the Capella HPC cluster
Thinking
Default: adaptive. The model decides for each request whether and how long to think, usually briefly.
reasoning_effort, as a request parameter or as chat_template_kwargs: {"reasoning_effort": "low"}:
low, medium, high, xhigh, max. low clearly shortens the thinking, max lengthens it somewhat,
and medium, high and xhigh stay close to the default. Other values (none, minimal, numbers)
are accepted and behave like the default.
Numeric effort: no.
Turn off: not possible. reasoning_effort: "low" comes closest; simple questions are then answered without
thinking. chat_template_kwargs: {"thinking_mode": ...} (MiniMax-M3) has no effect on this model.
Tool calling
Yes
Limits
At most 4 images per request. Video input is not supported.
The whole model shares a cache of about 1.1M tokens, so one request near the full 1M context fills it almost completely.
Several very long requests are processed one after another.
Large images use many tokens: a 1024 × 1024 image about 1,400, a 3584 × 3584 image about 5,400.