MiniMax M3.1 Flash 上线：速度提升 3 倍，价格降低 70%

稀宇科技推出新一代大语言模型 MiniMax M3.1 Flash1。该模型相比前代推理速度提升近 3 倍，使用成本降低 70%1。M3.1 Flash 已在 MiniMax API、MPlan 和 MiniMax Code 中上线1。

在定价方面，M3.1 Flash 的输入价格为每百万 Token 1 元，输出价格为每百万 Token 4 元，缓存读取价格为每百万 Token 0.1 元1。模型在多个评测中展现出强劲实力，在 DeepSWE v1.1 评测中取得 73.7% 的成绩1；在 Agent 任务方面，OSWorld V2 中得分为 61.7 分，Agents' Last Exam (ALE) 中得分为 29.6 分1；在网络安全领域的 CyberGym 基准测试中取得 82% 的高分1。

MiniMax 计划于稍后开放 M3.1 Flash 的权重，Pro 版本也会随后推出1。此前，Space Bunny 作为 M3.1 Flash 的早期版本，已处理了超过 200T Token 的数据1。

XiaoYu Technology has unveiled MiniMax M3.1 Flash, the latest generation large language model designed to deliver substantial improvements across Agent tasks, programming, Computer Use, and cybersecurity applications 1. The new model achieves a nearly threefold increase in inference speed while reducing operational costs by 70% compared to its predecessor 1. The model is now available on MiniMax API, MPlan, and MiniMax Code platforms 1.

Pricing for M3.1 Flash has been set at 1 yuan per million input tokens, 4 yuan per million output tokens, and 0.1 yuan per million tokens for cache reads 1. Across multiple benchmarks, the model demonstrates competitive performance: it achieved 73.7% on DeepSWE v1.1, scored 61.7 on OSWorld V2, and reached 29.6 on Agents' Last Exam (ALE) 1. In cybersecurity testing, M3.1 Flash attained a high score of 82% on the CyberGym benchmark 1.

The company notes that Space Bunny, an earlier version of M3.1 Flash, processed over 200 trillion tokens during development 1. XiaoYu Technology plans to release the model weights at a later date, with a Pro version to follow 1.