MiniMax M3.1 Flash bets on speed for AI agents

MiniMax has released M3.1 Flash, a new AI model designed for tasks that require software tools, web interfaces, and long sequences of actions. In an official blog post, MiniMax announces that the model is available through its API, M Plan subscription, and MiniMax Code development environment.

The company positions M3.1 Flash as a model for AI agents, meaning systems that can plan tasks and use applications or tools to complete them. MiniMax says the model improves performance in business automation, software development, computer use, and cybersecurity while lowering overall agent usage costs by 70 percent compared with its predecessor.

Benchmark claims and speed

MiniMax reports that M3.1 Flash scores 73.7 percent on DeepSWE v1.1, a benchmark for software engineering tasks such as understanding code, adding features, fixing bugs, and testing results. It also reports a score of 61.7 on OSWorld V2, which evaluates how well AI systems operate computers and work across applications.

For business workflows, the company cites a score of 76.2 on BankerToolBench. That benchmark covers multi step planning, document consistency, financial compliance, and connected tool calls. MiniMax also reports an 82 percent result on CyberGym, which measures security tasks including finding and validating vulnerabilities.

The company says the model can exceed 200 tokens per second in output speed, with median speeds above 170 tokens per second. It attributes the improvement partly to more aggressive sparsity, a model design approach intended to reduce the computing resources needed during inference.

Pricing and upcoming weights

MiniMax lists flat prices of $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 per million cached input tokens. The company says it plans to release the model weights on October 23, allowing others to run or adapt the model outside MiniMax’s hosted services.