MiniMax has officially launched its new M3.1 Flash model, which previously operated under the alias Space Bunny and processed over 200T tokens. The updated model delivers a threefold increase in inference speed and a 70% reduction in usage costs compared to its predecessor.
The model demonstrates high performance in complex agentic tasks, software development, and cybersecurity. In the DeepSWE v1.1 benchmark, it achieved a score of 73.7%. It also leads among domestic models in the OSWorld V2 test with 61.7 points and the Agents' Last Exam (ALE) with 29.6 points.
In the CyberGym benchmark, the model scored 82%, showing capabilities in vulnerability discovery, code auditing, and penetration testing. It is designed to handle multi-step planning and tool-chain calls, scoring 76.2 points in the BankerToolBench evaluation.
The pricing for the M3.1 Flash is set at ¥1 per million tokens for input and ¥4 per million tokens for output, with cache reads priced at ¥0.1 per million tokens. The model is currently available via the MiniMax API, MPlan, and MiniMax Code platforms.
MiniMax confirmed that the M3.1 Flash is part of its focus on balancing efficiency and performance. The company also announced plans to release a Pro version of the model in the future, which will prioritize higher intelligence capabilities.
