MiniMax Releases M3.1 Flash: 3x Performance Boost, 70% Cost Reduction
On the evening of October 19, MiniMax officially released its next-generation Flash model, M3.1 Flash. This model achieves significant improvements in complex Agent tasks, computer operations, software development, and cybersecurity, with online inference speed increasing by nearly 3 times and usage costs reduced by 70%. M3.1 Flash performed excellently in benchmarks such as AutomationBench-AA and DeepSWE v1.1, with API prices significantly lowered; input and output unit prices decreased by more than 50%, aiming to support complex workflows at a lower cost.
On the evening of October 19, MiniMax officially launched its next-generation Flash model, M3.1 Flash. Previously unveiled under the anonymous identity 'Space Bunny', this release marks a comprehensive performance upgrade. According to official introductions, M3.1 Flash shows significantly enhanced performance in complex tasks involving multi-step planning, continuous tool invocation, and long-horizon execution, while achieving dual breakthroughs of nearly 3 times faster online inference speed and a 70% reduction in usage costs.
In automation and complex task processing, M3.1 Flash ranked highly in the AutomationBench-AA benchmark, capable of fully solving real-world enterprise business workflow problems in sales, marketing, operations, customer service, finance, and HR. Additionally, it scored 76.2 in the BankerToolBench test, demonstrating strong capabilities in multi-step planning, cross-document consistency maintenance, financial compliance, and chained tool invocation.
Programming and software engineering capabilities are another highlight of this model. In the DeepSWE v1.1 evaluation, M3.1 Flash achieved a score of 73.7%, showing excellent long-chain task execution capabilities and the ability to support complex software development workflows at a lower cost. This performance indicates high reliability and efficiency when handling high-complexity code tasks.
Computer operations and professional task execution are also key areas of this update. M3.1 Flash scored 61.7 in OSWorld V2, which measures AI computer operation capabilities, and 29.6 in Agents' Last Exam (ALE), which assesses real professional work scenarios. Both metrics rank first among domestic models, marking that the model has reached industry-leading levels in simulating human operations and completing specific professional tasks.
Breakthrough progress was also made in the cybersecurity field. M3.1 Flash achieved a high score of 82% in the CyberGym benchmark, demonstrating leading vulnerability discovery and verification capabilities. After in-depth evaluation by internal security experts, the model performed excellently in scenarios such as code auditing, vulnerability patching, penetration testing, malicious code analysis, and reverse engineering, supporting enterprise complex security system construction with low cost and high efficiency.
Regarding inference efficiency optimization, M3.1 Flash adopts a more aggressive sparse attention architecture. In daily usage scenarios, its output speed can reach over 200 TPS, with the median (P50) stably reaching over 170 TPS. For Agentic-type tasks, the model's comprehensive usage cost is only one-third of the previous generation, greatly enhancing economic viability for large-scale applications.
Accompanying the performance improvement, MiniMax simultaneously lowered API call prices. The input cost for M3.1 Flash is 1 RMB per million tokens, the output cost is 4 RMB, and the cache read cost is 0.1 RMB. Compared to the previous generation M3 model, both input and output unit prices have decreased by more than 50%, and the cache read unit price has decreased by over 70%. It also supports a context window of up to 512K tokens, further lowering the barrier to use for developers.
The release of M3.1 Flash intensifies competition among domestic Flash models in the field of complex tasks. Two days prior, DeepSeek updated V4.1 Flash-1017, focusing on improving project development capabilities. In comparison, DeepSeek scores higher in software engineering evaluations, while MiniMax M3.1 Flash has advantages in cross-SaaS business flows and cybersecurity. Together, they form a complementary product portfolio, jointly driving domestic models to approach the performance levels of top closed-source models while maintaining low prices.
KEY HIGHLIGHTS
Key highlights
MiniMax released M3.1 Flash on October 19; the previously anonymous version 'Space Bunny' was its early form.
The model excels in complex Agent tasks, cross-SaaS business flows, and cybersecurity fields, ranking first among domestic models in multiple evaluations.
Adopting an aggressive sparse attention architecture, it achieves an output speed of over 200 TPS in daily scenarios, reducing Agentic task costs to one-third of the previous generation.
API pricing has been reduced: 1 RMB per million tokens for input and 4 RMB for output, representing a price reduction of over 50% compared to the previous generation M3.
Forms a complementary competitive relationship with DeepSeek V4.1 Flash, jointly driving the evolution of domestic Flash models into mainstay roles for complex tasks.