MiniMax launches M3.1 Flash, saying it delivers "a generational leap" in agentic tasks, coding, computer use, and cybersecurity, with weights due October 23

Agentic task costs fall by 70%, making complex AI-driven workflows significantly cheaper to run.

MiniMax has launched M3.1 Flash, an AI model that claims to offer significant improvements in agentic tasks, coding, computer use, and cybersecurity. The company states that M3.1 Flash represents a "generational leap" in performance, with inference speeds nearly three times faster and usage costs reduced by 70% compared to its predecessor. The model demonstrates strong performance across various benchmarks, including AutomationBench-AA for complex enterprise workflows, BankerToolBench for financial compliance and multi-step planning, and DeepSWE v1.1 for software development. It also shows proficiency in OSWorld V2 and Agents’ Last Exam (ALE) for computer operation and real-world task execution, achieving top scores among Chinese models in these areas. In cybersecurity, M3.1 Flash scored 82% on CyberGym and performed well in vulnerability discovery, auditing, and penetration testing. The model's inference efficiency is enhanced by aggressive sparsity, achieving over 200 tokens per second in everyday use. M3.1 Flash is available via the MiniMax API, M Plan, and MiniMax Code, with model weights to be released on October 23.

MiniMax's M3.1 Flash positions itself as a cost-effective alternative for advanced AI applications, directly challenging established models in agentic tasks, coding, and cybersecurity. By prioritizing inference efficiency and cost reduction, the company aims to democratize access to powerful AI capabilities for businesses that previously found them prohibitively expensive. This move signals a potential market shift towards more specialized, cost-optimized models designed for specific enterprise functions.

The emphasis on a 70% cost reduction for agentic tasks, coupled with performance parity with leading models on benchmarks like AutomationBench-AA and DeepSWE, suggests that the era of prohibitively expensive AI-powered automation may be drawing to a close. Businesses leveraging AI for operations, customer service, or software development should anticipate a more competitive pricing landscape and evaluate whether M3.1 Flash or similar offerings can streamline their workflows more economically.