ME AI Message: MiniMax officially launched the large model MiniMax M3 today. M3 is the only open-source large model currently equipped with the three cutting-edge capabilities—programming, ultra-long context, and native multimodality (model weights planned for open-sourcing within 10 days). It achieves international leading performance in code generation, agents, and desktop control, and is available for experience through MiniMax Code, Token Plan, and API.
For ultra-long contexts, M3 introduces the pioneering sparse attention architecture, MSA, which accelerates memory access by 4x over Flash-Sparse-Attention through KV block aggregation and query matching. Under a 1-million-token context, the new architecture reduces per-token computation to one-twentieth of the previous generation, achieving 9x faster prefill and 15x faster decoding.
On SWE-Bench Pro, M3 achieved a score of 59.0%, surpassing GPT-5.5 and Gemini 3.1 Pro, and approaching Opus 4.7. On Terminal Bench 2.1, it scored 66.0%. The team also developed an interactive user simulator framework to train the model’s long-term collaborative planning and multi-turn interaction capabilities.
In real-world testing, M3 autonomously ran for 12 hours to reproduce an ICLR 2025 paper and observed the DPO effect. In the Hopper-optimized FP8 operator task, it autonomously invoked tools 1,959 times within 24 hours, increasing hardware utilization from 7.6% to 71.3%, achieving a 9.4x speedup. In the PostTrainBench autonomous model training test, the model scored 0.37, slightly below Opus 4.7 and GPT-5.5, but clearly outperforming all other models.
MiniMax Code has been updated to support desktop control via computer use. The Plus plan, at 49 yuan per month, provides 6 billion tokens—approximately five times the capacity of Claude Pro. The API is now live, offering both thinking (reasoning) and non-thinking (fast) modes, with weights planned for open-source release within 10 days.(Source: MiniMax)