MiniMax Releases M3.1 Flash Model With 70% Lower Agent Costs

MiniMax Group Inc. has released M3.1 Flash, an updated large language model designed to lower the expenses associated with running AI agents. According to SiliconANGLE, this new version aims to make automated tasks more affordable by reducing the number of model calls required for complex jobs. The company states that agents now cost 70% less to operate than they did on the prior M3 model launched in June.

In plain English

AI agents are software programs that use large language models to perform multi-step tasks, such as coding or analyzing data. These agents often need to query the underlying model dozens of times to complete a single assignment. Each query, or token, incurs a cost. By making the model faster and cheaper per token, developers can run more queries without exceeding their budget. This allows for more complex automation without a proportional increase in spending.

The background

The M3.1 Flash model contains 428 billion parameters, a size nearly identical to the original M3 release. SiliconANGLE reports that MiniMax achieved higher speed through more aggressive sparsity during inference, meaning the model skips unnecessary calculations. Output speeds can reach 200 tokens per second at peak times. The model accepts text, images, and video, with prompts supporting up to 1 million tokens. It also features reasoning capabilities with five adjustable effort levels.

What changes now

MiniMax claims M3.1 Flash leads Chinese models in operating a computer, scoring 61.7 on the OSWorld V2 benchmark. It achieved 76.2 on BankerToolBench for banking workflows and 82% on CyberGym, a security benchmark built by University of California, Berkeley researchers. SiliconANGLE notes that CyberGym tests if AI can reproduce real software vulnerabilities. The model is priced at 15 cents per million input tokens, matching DeepSeek’s V4.1-Flash off-peak rate.

Key terms explained

Inference is the process where an AI model generates an output based on input data. Sparsity is a technique that allows the model to ignore less important data points during inference, speeding up processing and reducing computational load. AI agents are autonomous programs that use models to plan and execute multi-step tasks. Teams usually monitor inference costs by tracking token usage and latency. Check whether your current workflows require high-effort reasoning or if lower levels suffice for faster, cheaper results.

What to do and how to stay safe: MiniMax

• Review your current AI agent workflows to determine if switching to a cheaper model like M3.1 Flash reduces operational costs without sacrificing accuracy.
• Verify the security benchmarks cited by vendors, as independent validation of scores on tests like CyberGym may not yet be available for all claims.
• Monitor token usage rates closely when deploying new models, as pricing tiers and off-peak discounts can significantly impact total expenditure over time.
• Test the model’s reasoning levels to find the optimal balance between speed and capability for your specific application requirements.

Step-by-step guide: Stop AI Chatbots Leaking Sensitive Business Data

General security guidance from the Vector Update newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

How much does M3.1 Flash cost compared to DeepSeek?

M3.1 Flash costs 15 cents per million input tokens, which matches the off-peak rate of DeepSeek’s V4.1-Flash model according to SiliconANGLE.

Can M3.1 Flash handle security-related tasks?

Yes, MiniMax states the model can perform code auditing, penetration testing, and malware analysis, scoring 82% on the CyberGym security benchmark.

Is M3.1 Flash faster than the previous M3 model?

SiliconANGLE reports that MiniMax claims inference is nearly three times as fast, with output topping 200 tokens per second at everyday peak times.