by MiniMaxCN
Deliver strong coding and agent skills in a model that uses only a small slice of its size per step, keeping it cheap and fast to run so capable agents are affordable at scale.
M3 out since June 2026 with downloadable weights, succeeding M2.5; reads text, images and video.
Updated 1 Jun 2026Checked 11 Oct0 updates this week
A mixture-of-experts design runs about 23B of its ~428B parameters for each token, so it costs far less to run than its size suggests.
MiniMax Sparse Attention lets each token look at only part of the earlier text: 9x faster reading and 15x faster writing than M2 at 1M tokens.
The base model was trained using reinforcement learning across real-world task distributions to optimize for coding, agentic behavior and problem solving.
M3 takes images and video as input directly, not through a separate add-on model, and answers in text.
| Spec | MiniMax M3 | MiniMax M2.5 | Change |
|---|---|---|---|
| Total parameters | ~428 billionR (reported) | 230 billion (M2.5) | —no comparable change |
| Active parameters per token | ~23 billionR (reported) | 10 billion (M2.5) | —no comparable change |
| Context window | 1M tokensR (reported) | — | —no comparable change |
| SWE-Bench Verified score | 80.5%R (reported) | 80.2% (M2.5) | ▲ 0.3 |
| MMMU Pro (image reasoning) | 78.1%R (reported) | — | —no comparable change |
| Weights licence | MiniMax community licenceR (reported) | MIT (M2.5) | —no comparable change |
R reported by the company
MiniMax
M3 428B MoE model, open weights, Hailuo AI, Talkie
Shanghai-based AI company founded in December 2021 by computer vision researchers from SenseTime. Develops multimodal models and consumer apps (Talkie, Hailuo AI). Listed on Hong Kong Stock Exchange in January 2026.
