TL;DR
An official release of MiniMax-M3, a natively multimodal sparse model with about 428B total and 23B active parameters, a one-million-token context window, and agent-oriented coding and cowork capabilities.
Why it matters
M3 joins long-context architecture and RL-style capability scaling. Sparse attention lowers the cost of persistent agent context, while MaxProof shows the same model family using generative-verifier RL and population-level test-time search for formal reasoning.
Key findings
- 01
The official model card describes native text, image, and video training from the first step, with roughly 428B total parameters and 23B activated per token.
- 02
The MSA technical report reports 28.4× lower attention compute at one-million-token context on a 109B experimental model, plus 14.2× prefill and 7.6× decode speedups on H800 hardware.
- 03
The companion MaxProof work trains proof generation, verification, and critique-conditioned repair with generative-verifier RL, then scales inference through population search and tournament selection.
Scaling dimensions
Models, methods & benchmarks
- Models
- MiniMax-M3
- Algorithms
- MiniMax Sparse Attention, Generative-verifier RL, MaxProof, Population-level test-time scaling
- Benchmarks
- IMO 2025, USAMO 2026, Long-horizon agent benchmarks
Topics