China's large model competition has seen a significant technological breakthrough. On June 1st, Xiyu Technology officially launched its new generation of large language model - MiniMax M3. With cutting-edge programming capabilities, the ability to support an ultra-long context window of up to 1M (1 million), and native multimodal interaction capabilities, it becomes the first model in China to integrate these three top-tier technical indicators, and is also currently the only open-source model in the world with such performance.

In multiple authoritative evaluations that measure the core competitiveness of large models, MiniMax M3 delivered outstanding results. In the challenging programming evaluation set SWE-Bench Pro, M3 achieved an excellent score of 59.0%, which not only surpassed GPT-5.5 and Gemini3.1Pro, but also approached the industry benchmark Claude3.5Opus (corresponding to Opus4.7 in the original text); while in the Claw-Eval evaluation testing AI agent scheduling capabilities, and the OmniDocBench test examining multimodal document parsing capabilities, M3 achieved the highest industry scores or exceeded Gemini3.1Pro.

image.png

M3 can maintain efficient operation while achieving a performance boost, mainly due to major innovations in its underlying architecture. The model adopts a new sparse attention architecture called MSA (MiniMax Sparse Attention). With this underlying logic, when processing the ultra-long context of 1 million, the computational cost per Token is reduced to one-tenth of the previous generation model. This directly leads to a significant speed improvement: during the Prefilling (pre-filling) stage, the model runs more than nine times faster; and during the Decoding (decoding generation) stage, the acceleration is more than 15 times, significantly reducing computing costs.

As a native multimodal model, M3 not only deeply understands text, but also supports active input of images and videos, and even directly completes high-level computer desktop automation operations. Along with the release of the model, MiniMax also upgraded its AI programming assistant product MiniMax Code, and introduced a new Token Plan subscription scheme for developers, available in three tiers: Plus (49 yuan/month), Max (119 yuan/month), and Ultra (469 yuan/month), to meet the needs of different development scales.

Currently, the API interface of M3 is officially open for public use. To attract more developers to join the ecosystem, the official also launched a limited-time offer, with the 512k context version offering a 50% discount for seven days. In addition, MiniMax officially promised that the complete weights of the model and detailed technical reports will be officially open-sourced to global developers within 10 days. This "full disclosure" open-source confidence will undoubtedly further accelerate the explosion of AI innovation at the edge and application layers.