The official claims that this model achieves generational improvements in agentic tasks, programming, computer use, and cybersecurity, with online inference speed increased by nearly 3x and usage costs reduced by 70%.Author and source: AIBase
On October 19, MiniMax officially launched its next-generation Flash model, M3.1Flash, which evolved from the early version used anonymously to process over 200T tokens under the name Space Bunny. The company states that this model delivers generational improvements in agentic tasks, programming, computer use, and cybersecurity, with online inference speeds increased by nearly 3x and usage costs reduced by 70%.
On benchmark performance, M3.1Flash ranks among the top on AutomationBench-AA, capable of executing real business workflows across SaaS systems for sales, marketing, operations, customer service, finance, and HR. BankerToolBench scores 76.2 and excels in multi-step planning, cross-document consistency, financial compliance, and chained tool invocation.
In terms of programming, DeepSWE v1.1 achieves 73.7%, reaching the level of top-tier models. This evaluation requires the model to independently complete code comprehension, feature development, bug fixing, and test validation, highlighting its capability in executing long-chain tasks.
In terms of computer usage, OSWorld V2 scored 61.7, and Agents' Last Exam scored 29.6, both ranking first among domestic models. In cybersecurity, OSWorld V2 achieved 82% on the CyberGym benchmark, excelling in scenarios such as code auditing, vulnerability patching, penetration testing, malicious code analysis, and reverse engineering.
In terms of inference efficiency, the model employs a more aggressive sparse attention architecture, achieving output speeds of over 200 TPS in everyday scenarios, with a stable median P50 of over 170 TPS. Regarding pricing, input costs RMB 1 per million tokens, output costs RMB 4 per million tokens, and cached reads cost RMB 0.1 per million tokens. The overall cost for Agentic-type tasks is reduced by 70% compared to the previous generation.
M3.1 Flash is now live with MiniMax API, MPlan, and MiniMax Code. Weights will be released shortly. The official team describes this as an ongoing exploration of the balance between efficiency and capability, with a Pro version aimed at pushing the limits of intelligence to follow soon. It sets a new benchmark for performance versus cost, continuing Flash-series models’ product strategy of trading efficiency for enhanced capability.