Inference model · MINIMAX

MiniMax M2.7

1M-token context, 22B active. The model you reach for when nothing else fits.

Request MiniMax M2.7 access Back to catalog

MiniMax M2.7 is the frontier MoE option in our catalog. 1M-token context window, 22B active parameters, suitable for the workflows that don’t fit anywhere else: whole-codebase reasoning, multi-hour transcript analysis, very-long-horizon agents.

We’re holding this model in “pending license review” status at founding launch because the upstream license terms are still being clarified for commercial hosting. We’ll confirm before founding plans are charged.

When to pick it

  • Whole-codebase coding agents (1M context fits most production repos)
  • Multi-document research over 100+ sources in a single context
  • Long-horizon agent runs where you can’t tolerate summarization losses

When to look elsewhere

  • Anything that fits in 256K → use Qwen 3.5 122B A10B instead, cheaper
  • Latency-sensitive workflows — MiniMax M2.7 trades latency for context

License under review at founding launch.

We're confirming the commercial-hosting license terms with the model authors before turning this on for founding customers. If the license terms don't allow our usage model, we'll offer founding customers their deposit back or a credit toward an alternative model — your choice.