MiniMax Releases M3.1 Flash, Outperforming Previous Model in Cybersecurity and AI Tasks

MiniMax’ M3.1 Flash is a next-generation Flash model that delivers a ‘generation leap’ in agent tasks, coding, computer usage, and cybersecurity, according to the company. The model achieves nearly three times faster online inference and reduces costs by 70% compared to its predecessor. Space Bunny, an anonymous model that processed over 200 billion tokens during its stealth run, was a precursor to M3.1 Flash.

OpenRouter, the platform that hosted Space Bunny, reported that the model topped global rankings during its anonymous test period, handling 56.7 billion tokens and achieving a 38% improvement in token processing time, a 45% increase in throughput, and a 50% reduction in error rate. MiniMax attributed the model’s success to user feedback during the anonymous test period, which helped identify and fix issues.

MiniMax’ M3.1 Flash outperforms its predecessor in various benchmarks, including AutomationBench-AA, DeepSWE, and CyberGym, achieving scores of 72.5%, 73.67%, and 82%, respectively. The model also excels in computer usage, software development, and cybersecurity tasks, with scores of 61.7, 29.6, and 82%, respectively.