MiniMax M3.1 Flash launched on October 19, 2026, with the company claiming nearly threefold faster online inference and 70% lower overall usage costs for agentic tasks than its predecessor. The model is available through the MiniMax API, M Plan and MiniMax Code, targeting software development and multistep enterprise workflows.
Key Takeaways
- MiniMax says the model combines faster inference with lower costs for agentic tasks.
- The release targets coding, cross-application automation and cybersecurity work.
- Hosted access is available now, while model weights are scheduled for a later release.
Table of Contents
- MiniMax M3.1 Flash targets cheaper task execution
- Coding and enterprise automation define the product pitch
- Hosted access precedes the model-weight release
- Conclusion
MiniMax M3.1 Flash targets cheaper task execution
MiniMax’s M3.1 Flash launch announcement puts efficiency at the center of the release. The company says it improved inference through more aggressive sparsity, with output speeds reaching more than 200 tokens per second in everyday use and median speeds exceeding 170 tokens per second.
The 70% cost-reduction claim applies to overall usage costs for agentic tasks compared with the preceding model. It is a company-reported comparison, not a guarantee that every customer’s bill will fall by the same amount.
For developers, the distinction is practical: the listed token prices establish usage charges, while the cost of completing a workflow also depends on how much input and output the task requires.
| Item | Details |
|---|---|
| Input tokens | $0.15 per 1 million tokens |
| Output tokens | $0.60 per 1 million tokens |
| Cached input tokens | $0.015 per 1 million tokens |
Coding and enterprise automation define the product pitch
MiniMax says M3.1 Flash can handle workflows across software-as-a-service systems, including sales, marketing, operations, customer service, finance and human resources. These tasks involve multiple steps and tool calls rather than a single generated response.
The company reported a 73.7% score on DeepSWE v1.1, which tests software-development tasks such as understanding code, adding features, fixing problems and validating work. It also reported 61.7 points on OSWorld V2, a benchmark for computer operation and cross-application execution.
Those benchmark results describe performance in specified evaluations. They do not establish customer adoption, revenue gains or equivalent results across every enterprise environment.
Hosted access precedes the model-weight release
The launch makes the model available through three existing MiniMax channels: its API, M Plan and MiniMax Code. MiniMax said model weights would follow later but did not provide a release date.
The company also said a Pro model focused on greater capability would arrive later. No timetable was supplied for that release.
Conclusion
The launch gives developers immediate hosted access to MiniMax’s lower-cost model offering. The next announced step is the release of model weights, although MiniMax has not set a date.
Read More News
Disclaimer
This FinanceVelix article provides general information, not personalized financial, investment, legal, or tax advice. Market prices, interest rates and economic conditions can change; investments carry a risk of loss. Facts reflect the material available when published and may change. Check important information independently and consult a qualified professional when needed. Mentioning an asset, company, or service is not an endorsement.
