428B M3.1 Flash: 3x Faster, 70% Cheaper Agents, 82% CyberGym Score
The new 428B parameter model offers three times faster speed and lower costs for agentic tasks, challenging competitors in the AI market.
Key points
- M3.1 Flash reduces agentic task costs by 70% compared to the previous M3 model.
- Inference speeds are approximately three times faster, reaching 200 tokens per second.
- The model scores 82% on CyberGym, a benchmark for reproducing software vulnerabilities.
Chinese developer MiniMax Group Inc. released M3.1 Flash, an update to its large language model that significantly lowers the expense of running AI agents. According to SiliconANGLE, this release targets the high costs associated with agents that require multiple model calls to complete single jobs.
The risk
The primary shift here is economic rather than technical vulnerability. By making agent operations cheaper and faster, MiniMax lowers the barrier for deploying automated systems at scale. This could increase the volume of automated interactions on the internet, potentially amplifying the impact of any malicious agents if security controls are not robust.
Who is exposed
Software developers and organizations planning to deploy AI agents for coding, cybersecurity, or banking workflows are the direct audience for this model. As Vector Tribune reported on October 19, MiniMax previously highlighted cybersecurity capabilities for this release. The model is designed for developers who need to balance efficiency with complex task execution.
Cost and compliance
MiniMax charges a flat rate of 15 cents per million input tokens for M3.1 Flash. This pricing directly competes with Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co. Ltd.’s V4.1-Flash model, which charges the same rate off-peak. While DeepSeek’s model resolved 77% of tasks on the DeepSWE v1.1 test, M3.1 Flash managed 73.7%, according to SiliconANGLE. Neither result has been independently verified.
What should you do?
Teams evaluating this model should verify the independent validity of the performance claims before integrating it into production environments. Developers should assess whether the cost savings justify the potential risks of using a model with unverified benchmark results for critical security or financial tasks.
The background in plain language
AI agents are software programs that use large language models to perform complex tasks by breaking them into smaller steps. Each step often requires a separate call to the model, which accumulates costs and time. Efficiency improvements in these models allow agents to complete more steps faster and cheaper. Teams usually monitor the trade-off between speed and accuracy when adopting new models, ensuring that cost reductions do not compromise the reliability of automated decisions or security outcomes.
Background: AI agents
AI agents are software systems that perceive their environment, make decisions, and execute actions to achieve specific goals. Unlike standard models, they maintain internal state and interact with external tools, creating unique security risks around tool execution, memory retention, and autonomous decision loops that require specialized controls.
Read the full guide: AI Agents Explained: Answers to Security and Architecture Questions
What to do and how to stay safe: MiniMax
- Evaluate the specific cost structure of your current AI agent deployments against the new pricing to determine potential savings.
- Review the benchmark scores provided by the vendor and seek independent validation before trusting the model for critical security tasks.
- Monitor the model's performance in your specific use cases, particularly for code auditing and penetration testing, to ensure reliability.
- Stay updated on any independent audits or community feedback regarding the model's claims to inform your deployment decisions.
Step-by-step guide: AI Governance Explained: Building Rules for Autonomous Systems
General security guidance from the Vector Tribune newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
How does M3.1 Flash compare to DeepSeek's V4.1-Flash in terms of cost?
Both models charge a flat rate of 15 cents per million input tokens, according to SiliconANGLE.
What benchmarks did M3.1 Flash achieve in cybersecurity testing?
The model scored 82% on CyberGym, a benchmark built by University of California, Berkeley researchers to test vulnerability reproduction.
Was the identity of the Space Bunny model revealed with this launch?
Yes, the launch settled that MiniMax was behind the anonymous Space Bunny model that appeared on OpenRouter Inc.’s service.
