MiniMax M3.1 Flash release notes

MiniMax has officially released M3.1 Flash, a next-generation model optimized for agentic workflows, software engineering, and computer operation. The model, previously tested under the internal codename "Space Bunny," aims to balance high-performance capabilities with aggressive pricing and inference efficiency.

Performance and Capabilities

M3.1 Flash is positioned as a specialized tool for complex, multi-step automation. According to MiniMax, the model achieves competitive scores on several industry benchmarks:

• Automation and Planning: It scores 76.2 on BankerToolBench, demonstrating proficiency in multi-step planning, financial compliance, and chained tool execution.
• Software Development: The model achieves 73.7% on DeepSWE v1.1, a benchmark focused on autonomous feature development and bug fixing.
• Computer Use: It records 61.7 on OSWorld V2 and 29.6 on Agents’ Last Exam (ALE), which the company claims is the highest among Chinese models for cross-application task execution.
• Cybersecurity: It reports an 82% score on CyberGym, with capabilities spanning code auditing, vulnerability patching, and penetration testing.

MiniMax attributes these gains to more aggressive sparsity in the model architecture, which enables output speeds exceeding 200 tokens per second (TPS) in optimal conditions, with median speeds of 170+ TPS.

Pricing and Availability

MiniMax has implemented a significant reduction in usage costs, claiming a 70% decrease compared to its predecessor, the M3. The pricing structure is set at:

• Input: $0.15 per 1M tokens
• Output: $0.60 per 1M tokens
• Cached Input: $0.015 per 1M tokens

The model is currently available via the MiniMax API, M Plan, and the MiniMax Code desktop application. The company has stated that model weights will be released at a later date.

Community Reception and Analysis

User reception in the developer community has been mixed, with significant skepticism regarding the reliance on benchmark scores. Many users noted that "Space Bunny" (the model's pre-release version) received varied feedback during testing, with some developers reporting that it struggled with coding tasks compared to competitors like DeepSeek V4.1 Flash or Qwen variants.

Critics have pointed to a trend of "benchmaxxing," where model performance on standardized tests may not translate to real-world utility. Some users noted discrepancies between MiniMax's internal benchmarks and their own independent testing. As one observer noted:

"The perverse incentives around benchmarks are now so bad that benchmark scores are effectively meaningless. Models can score highly and still be crap to use."

Others highlighted that while the cost-to-performance ratio is attractive, the actual "wall clock" time for task completion remains a critical factor that is often omitted from marketing materials. While some testers found the model effective for documentation and structural tasks, others cautioned that it did not consistently outperform existing alternatives in high-stakes coding environments.