MiniMax Launches M3.1 Flash: 3x Faster Inference at 70% Lower Cost

MiniMax M3.1 Flash Launches: 3x Speed Boost, 70% Price Reduction

On October 19, MiniMax announced the launch of its new model, MiniMax M3.1 Flash. The previously anonymous "Space Bunny," which processed over 200T tokens, was an early version of MiniMax M3.1 Flash. As a next-generation Flash model, M3.1 Flash delivers generational improvements in agentic tasks, coding, Computer Use, and cybersecurity capabilities. Meanwhile, online inference speed has increased by nearly 3x, and usage costs have dropped by 70%.

According to the introduction, M3.1 Flash has achieved comprehensive progress across multiple benchmarks, reaching leading-model levels in complex agent tasks (Cowork), Computer Use, software development (Coding), and cybersecurity.

"960x1221"

Complex Tasks and Automation Capabilities:

• Ranked at the top in AutomationBench-AA. It can fully solve real-world business problems across sales, marketing, operations, customer service, finance, and HR within actual enterprise workflows spanning multiple SaaS systems.
• Scored 76.2 on BankerToolBench, demonstrating proficiency in multi-step planning, cross-document/cross-file consistency, financial compliance, and chained tool calls for complex real-world scenarios.

Coding and Software Engineering Capabilities:

• Achieved a score of 73.7% on the DeepSWE v1.1 evaluation, reaching top-tier model performance. DeepSWE focuses on complex software tasks requiring models to independently handle code understanding, feature development, bug fixing, and test verification. M3.1 Flash exhibits excellent long-chain task execution capabilities, supporting complex software development workflows at lower costs.

Computer Use Capabilities:

• Scored 61.7 on OSWorld V2, which measures AI computer operation and cross-application task execution capabilities, and 29.6 on Agents' Last Exam (ALE), which evaluates computer operation skills in realistic, economically valuable scenarios. Both scores rank first among domestic models.
• M3.1 Flash demonstrates superior capabilities in real-environment interaction, tool usage, and long-chain task execution, moving beyond mere task understanding to actual operation and delivery.

Cybersecurity Capabilities:

• Achieved a high score of 82% on the CyberGym benchmark, showcasing leading vulnerability discovery and validation capabilities.
• In-depth evaluations by internal security experts indicate that the model performs excellently in scenarios such as code auditing, vulnerability patching, penetration testing, malicious code analysis, and reverse engineering, enabling low-cost, high-efficiency support for complex security infrastructure construction.

For agentic tasks, the overall usage cost of the model has decreased by 70% compared to previous generations. Input pricing is ¥1 per million tokens, output pricing is ¥4 per million tokens, and cache read pricing is ¥0.1 per million tokens.

"960x390"

Across three key benchmarks—Automation-Bench (measuring complex agent task execution), DeepSWE (measuring real-world coding capability), and Agents' Last Exam (ALE) (measuring computer operation in realistic, economically valuable scenarios)—M3.1 Flash achieves top-tier model performance at extremely low task costs, establishing a highly competitive new standard balancing performance and cost.

"960x688"

"960x720"

"960x720"

IT Home noted that M3.1 Flash is now officially available via the MiniMax API, MPlan, and MiniMax Code, with open weights expected to be released shortly. According to official statements, the Flash series represents our continuous exploration of balancing efficiency and capability, while the Pro version, which pursues the upper limits of intelligence, will also be available soon.

Reference