On June 1, MiniMax officially released its flagship large model M3. M3 uses the company’s self-developed new sparse attention architecture MSA (MiniMax Sparse Attention), supporting up to 1 million token ultra-long context. At this scale, the computation per token is only 1/20 of the previous generation model, with prefilling stage speed increased by over 9 times and decoding stage speed increased by over 15 times. In terms of programming capability, M3’s SWE-Bench Pro score is 59.0%, surpassing GPT-5.5 and Gemini 3.1 Pro, close to Claude Opus 4.7. It ranks first in the end-to-end Agent evaluation Claw-Eval and surpasses Opus 4.7 on SVG-Bench. M3 is also a native multimodal model, supporting image and video input, and has computer desktop operation (Computer Use) capabilities. MiniMax claims that M3 is the first domestic model to simultaneously possess cutting-edge Coding/Agentic capabilities, 1M ultra-long context, and native multimodal features, and it is currently the only open-source model that meets all the above conditions.
MiniMax disclosed three cases of long-term autonomous capabilities: M3 independently reproduced an ICLR 2025 best paper, running autonomously for about 12 hours, producing 18 commits and 23 experimental charts; optimized FP8 matrix multiplication operators on NVIDIA Hopper architecture GPUs, after 147 benchmark submissions and about 24 hours of continuous execution, pushing hardware peak utilization from 7.6% to 71.3%, achieving a 9.4x speedup; on PostTrainBench, it autonomously completed the full process of data synthesis → training → evaluation, scoring 0.37, close to GPT-5.5 (0.39) and Opus 4.7 (0.42). The Agent product MiniMax Code, released alongside M3, supports Agent Team concurrent collaboration and Computer Use. Token Plan is divided into three tiers: Plus ¥49/month (600 million tokens), Max ¥119/month (1.8 billion tokens), and Ultra ¥469/month (5.5 billion tokens). The API is now open, and the company plans to release a technical report and open-source model weights within the next 10 days.