IT Home On October 19, Xi Yu Technology announced that the MiniMax M3.1 Flash model is now available. The Space Bunny that previously processed over 200T Tokens under an anonymous identity is actually an early version of the MiniMax M3.1 Flash. As a new generation Flash model, M3.1 Flash has seen a generational improvement in Agentic tasks, programming, Computer Use, and network security capabilities. Meanwhile, the online reasoning speed has increased nearly 3 times, and the usage cost has been reduced by 70%.
According to the report, M3.1 Flash has made comprehensive improvements in multiple benchmark tests. It has reached the level of a leading model in areas such as complex Agent tasks (Cowork), Computer Use, Software Development (Coding), and Network Security.
Complex tasks and automation capabilities:
It ranks at the top in AutomationBench-AA. In the real business workflows across SaaS systems, it can fully address the actual issues in sales, marketing, operations, customer service, finance, and HR.
In BankerToolBench, it gets 76.2 points. It is proficient in multi-step planning, consistency across documents/files, financial business compliance, and tool chain calls for real, complex issues.
Programming and Software Engineering Skills:
In the DeepSWE v1.1 evaluation, it achieved a score of 73.7%, reaching the level of a top model. DeepSWE focuses on complex software tasks, requiring the model to independently understand code, develop features, fix issues, and verify testing. M3.1 Flash demonstrates excellent capability in executing long-chain tasks, capable of supporting complex software development workflows at lower costs.
Computer Use Ability:
In OSWorld V2, which measures the operation capabilities of AI computers and the execution of cross-app tasks, it got 61.7 points, and in Agents' Last Exam (ALE), which measures computer operation capabilities in real and economically valuable scenarios, it got 29.6 points. Both scores ranked first among domestic models.
M3.1 Flash demonstrates excellent real-world environment interaction, tool usage, and long-chain task execution capabilities, allowing it to progress from understanding tasks to actual operation and delivery.
Cybersecurity capabilities:
Achieved a high score of 82% in the CyberGym benchmark test, demonstrating leading vulnerability discovery and verification capabilities.
After a deep evaluation by internal security experts, the model performs excellently in scenarios such as code auditing, vulnerability patching, penetration testing, malicious code analysis, and reverse engineering. It can support complex security system construction at low cost and with high efficiency.
In agentic class tasks, the comprehensive use cost of the model has been reduced by 70% compared to previous generations of models.The input price is 1 yuan per million Tokens, the output price is 4 yuan per million Tokens, and the cache read price is 0.1 yuan per million Tokens.。
In the Automation-Bench benchmark that measures the execution capability of complex Agent tasks, DeepSWE that measures real programming ability, and Agents' Last Exam (ALE) that measures computer operation ability in real-world, economically valuable scenarios, M3.1 Flash achieved top-model-level performance at extremely low task costs, establishing a highly competitive new standard between performance and cost.
IT Home notes that M3.1 Flash has now been officially launched in the MiniMax API, MPlan, and MiniMax Code, and the access will be opened later. According to the official introduction, Flash is our continuous exploration of the balance between efficiency and capability, and the Pro version that aims for the upper limit of intelligence will also be available later.
Reference
Ad notice: The text contains external link references (including but not limited to hyperlinks, QR codes, passwords, etc.) designed to provide more information, save selection time, and the results are for reference only. Articles on IT Home that include external links all include this notice.