The Chinese company MiniMax has officially unveiled its new flagship development, the M3 artificial intelligence model. This model is based on the new MiniMax Sparse Attention (MSA) architecture, designed for efficient processing of large volumes of data. This was reported by Ixbt.com .
The MiniMax M3 model supports an ultra-long context window of up to 1 million tokens. It is a multimodal system capable of analyzing not only text but also images and videos. It is also adapted for agentic scenarios, such as controlling a computer desktop and executing complex multi-step tasks.
According to the company's internal tests, M3 is recording high results in a number of benchmarks. Specifically, in the SWE-Bench Pro test, it outperformed GPT-5.5 and Gemini 3.1 Pro, approaching the level of Opus 4.7. In creating SVG graphics, it even surpassed Opus 4.7.
In the multimodal test OmniDocBench, the system proved its superiority over Gemini 3.1 Pro. In the Claw-Eval rating for agentic scenarios, the model is showing one of the best results. Currently, access to the new model is available via MiniMax Code, API, and special subscription plans.
Read “Zamin” on Telegram!
