MiniMax 发布 M3.1 Flash 模型

MiniMax 正式推出下一代 Flash 模型 M3.1 Flash，在 Agent 任务、代码编写、计算机使用和网络安全方面实现了代码级性能提升。1该模型在线推理速度相比前版本提升近三倍，使用成本下降 70%，输出速度可达 200+ TPS，中位数稳定在 170+ TPS。1定价方面，该模型按 $0.15/输入、$0.60/输出、$0.015/缓存输入（每百万 token）计费，已通过 MiniMax API、M Plan 和 MiniMax Code 提供。1

M3.1 Flash 在多个基准测试中与领先模型性能相当。1在 Agent 任务方面，该模型在 AutomationBench-AA 中排名靠前，可完成跨 SaaS 系统的真实企业工作流；在 BankerToolBench 得分 76.2，展现出多步规划和工具链调用能力。1代码编写方面，DeepSWE v1.1 基准测试中达成 73.7% 的成绩，可独立理解代码、开发功能、修复问题。1计算机使用领域，OSWorld V2 得分 61.7，Agents' Last Exam (ALE) 得分 29.6，在中文模型中均排名第一。1网络安全方面，该模型在 CyberGym 得分 82%，在漏洞发现和验证方面领先。1

Space Bunny 是 M3.1 Flash 的早期版本，已处理超过 200 万亿 token。1

MiniMax has unveiled M3.1 Flash, a next-generation AI model designed to deliver substantial improvements across multiple dimensions of language model performance and efficiency.1 The model achieves nearly threefold faster online inference speed while reducing operational costs by 70 percent.1

The M3.1 Flash demonstrates competitive capabilities across diverse benchmark assessments. In agent task execution, the model ranks prominently on AutomationBench-AA and can execute genuine enterprise workflows spanning multiple SaaS systems.1 On BankerToolBench, it achieves a score of 76.2, demonstrating proficiency in multi-step planning and tool chain invocation.1 For software engineering tasks, M3.1 Flash reaches 73.7 percent on DeepSWE v1.1, enabling independent code comprehension, feature development, and bug resolution.1 In web navigation benchmarks, the model scores 61.7 on OSWorld V2 and 29.6 on Agents' Last Exam (ALE), both rankings first among Chinese-language models.1 The model also achieves 82 percent on CyberGym, leading in vulnerability discovery and validation.1

From an operational perspective, M3.1 Flash delivers output speeds exceeding 200 tokens per second, with median (P50) performance consistently maintaining over 170 tokens per second.1 The model is priced at $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 per million cached input tokens.1 Access is available through MiniMax API, M Plan, and MiniMax Code platforms.1 The model builds upon Space Bunny, an earlier version that processed over 200 quadrillion tokens.1