MiniMax-M3-Free targets long context and lower agent costs

MiniMax-M3-Free targets long context and lower agent costs

MiniMax-M3-Free is an open-weight 428B MoE multimodal . It supports up to a , so it can handle very large amounts of text in one run.

It also includes native video understanding. The listed result is 59% on , with a claim that it can rival Claude and GPT-5.5 at about one-sixth of the cost.

It can be self-hosted with Ollama or used through an API on OpenRouter. The material also includes guidance for , , and deployment.

Key points

  • Open-weight 428B MoE model aimed at advanced AI work.
  • Supports a for very large inputs.
  • Includes native video understanding.
  • Claims 59% on and about one-sixth the cost of top .
  • Can be run with Ollama or accessed through .
Read original