Overview

MiniMax-M3 was trained on mixed text, image and video data from the first step. Its MiniMax Sparse Attention (MSA) operator reduces the cost of very long contexts: MiniMax reports 9x faster prefill and 15x faster decoding than M2 at 1M tokens. It offers enabled, adaptive and disabled reasoning modes.

Highlights

  • About 428B total, 23B active parameters
  • Native text, image and video input
  • 1M-token context window
  • Three reasoning modes, including adaptive
  • Technical report on MiniMax Sparse Attention

Typical uses

Coding and “cowork” agents, long-document analysis and multimodal assistants.

Things to consider

  • The MiniMax Community License permits non-commercial use; commercial use needs separate terms.
  • The weights were published after the API launch.

Specifications are taken from the model card on Hugging Face and the provider’s announcement. Benchmark results depend on evaluation settings; see how AI benchmarks work.