MiniMax releases M3 open-weight model
MiniMax has launched M3, which it says is the first open-weight model to combine near-frontier coding and agent performance, a 1M-token context window, and native image and video input.
MiniMax announced M3 today. It is available now through MiniMax Code, the company's Token Plan and its API. The company says the open weights will follow about 10 days after launch. MiniMax calls long context, native multimodality and desktop computer use "table stakes" for closed frontier models, and says M3 is the only open-weight model that offers all three.
The long context comes from MiniMax Sparse Attention (MSA), a new architecture that filters key-value blocks before computing attention, so cost does not grow quadratically with context length. MiniMax says MSA splits KV blocks more precisely than DSA and MoBA. Its kernel loops over KV blocks and gathers the queries that hit each one, which the company says makes it more than 4× faster than the open-source Flash-Sparse-Attention and flash-moba. At 1M tokens, MiniMax says M3 needs 1/20 the per-token compute of the previous generation, with prefill more than 9× faster and decoding more than 15× faster. In ablations, it says, MSA matched full attention on most capabilities.
MiniMax reported these benchmark scores:
- SWE-Bench Pro: 59.0%
- Terminal-Bench 2.1: 66.0%
- SWE-fficiency: 34.8%
- KernelBench Hard: 28.8%
- MCP Atlas: 74.2%
All of these numbers are self-reported. MiniMax describes M3 as "approaching" leading closed models on coding. On several tasks it still trails Opus 4.7 and GPT-5.5.
For training, MiniMax built a simulator that imitates developers working with the model over many turns. M3 was also trained on mixed image, video and text data from the first step, and MiniMax rebuilt its text pretraining pipeline to produce more of that interleaved data. In one internal test, M3 spent nearly 12 hours reproducing an ICLR 2025 Outstanding Paper on its own, producing 18 commits and 23 experimental figures.