MiniMax Group Inc. has launched M3.1 Flash, an updated large language model designed to make AI agents more cost-effective and efficient. The new model reduces operational expenses by 70% compared to its predecessor while delivering near triple the inference speed.
- AI agent inference costs cut by 70%, inference speed nearly tripled
- M3.1 Flash scored highest among Chinese models on multiple efficiency tests
- Supports multimodal input and extensive prompt sizes, suitable for complex tasks
What happened
MiniMax Group Inc., a Chinese AI developer, has introduced the M3.1 Flash large language model, an upgrade to its June-launched M3 version. This new model aims to reduce the cost and increase the speed of running AI agents, which typically require numerous model calls to complete a task. By leveraging a technique of more aggressive sparsity during inference, M3.1 Flash achieves inference speeds close to three times that of the previous version.
Despite minimal change in model size, which remains at 428 billion parameters, M3.1 Flash enables AI agents to process prompts up to 1 million tokens and accepts text, image, and video inputs. The model offers developers configurable reasoning intensity across five levels from Low to Max, enhancing task adaptability. MiniMax has made the model accessible through its API, subscription M Plan, and desktop application.
Why it matters
Reducing AI deployment costs is a critical factor in expanding the practical use of AI agents in enterprise and technical workflows. By cutting expenses by approximately 70%, MiniMax’s M3.1 Flash lowers barriers for businesses requiring extensive AI assistance in tasks such as code auditing, penetration testing, and malware analysis. The significant speed improvements facilitate faster real-time interactions, increasing model usability in dynamic environments.
Benchmark results reinforce the model’s strength in diverse applications, including operating computer systems (OSWorld V2 score of 61.7), banking workflows (BankerToolBench score of 76.2), and cybersecurity simulations (82% on CyberGym). These assessments highlight its broad capability across technical and security domains, potentially making it an attractive option for companies aiming to automate complex and sensitive operations.
What to watch next
MiniMax plans to release the Flash model's weights publicly in the future, enabling wider adoption and further innovation by the AI development community. Additionally, a more capable Pro version of the M3.1 model is in the pipeline, which may provide enhanced performance or new features tailored for high-demand use cases.
The competitive landscape includes Hangzhou DeepSeek’s V4.1-Flash model, which charges a similar price and recently reported slightly higher software engineering task resolution. Observers should monitor market reactions, Uptake by enterprise customers, and further independent benchmarking results to judge MiniMax’s position amid rapid advances in Chinese AI large language models.