MiniMax launches M3.1 Flash, touting faster agent tasks at lower cost

MiniMax has launched M3.1 Flash, a model it is positioning for agentic work, coding and computer use. The company says online inference is nearly three times faster than its predecessor and that usage costs for agentic tasks are 70% lower. Those are MiniMax’s claims; the company’s announcement does not provide an independent comparison.

MiniMax reports results across several benchmarks, including 76.2 on BankerToolBench and 73.7% on DeepSWE v1.1. It also says M3.1 Flash scored 61.7 on OSWorld V2 and 29.6 on Agents’ Last Exam, ranking first among Chinese models on both. The company describes the tests as covering workflows such as multi-step tool use, software engineering and computer operation.

The company also reports an 82% score on CyberGym and says its internal security experts evaluated the model on code auditing, vulnerability patching, penetration testing, malware analysis and reverse engineering. MiniMax says output speeds can reach more than 200 tokens per second, with median speeds above 170 tokens per second. The announcement does not include independent validation of those performance figures.

M3.1 Flash is available through the MiniMax API, M Plan and MiniMax Code. MiniMax lists prices per million tokens of 15 cents for input, 60 cents for output and 1.5 cents for cached input. The company says model weights will be released later, without giving a date.

MiniMax describes Flash as an effort to balance capability and efficiency, and says a more intelligence-focused Pro model is also due later. It did not provide a release timeline for Pro in the announcement.