MiniMax releases M3.1 Flash model, says it is faster and 70% cheaper for agent tasks
The company says the model matches leading systems on agent, coding and computer-use benchmarks. Its announcement gives no scores, and the claims have not been independently tested.
MiniMax has released a new AI model called M3.1 Flash, according to an announcement on the company's own research blog. This story rests on that one report, which comes from the company itself.
MiniMax said an early version of the model had already been running in public without the company's name on it, under the label "Space Bunny." The company said that early version processed more than 200 trillion tokens while operating anonymously.
According to MiniMax, M3.1 Flash delivers "a generational leap in agentic tasks, coding, Computer Use and cybersecurity." The company said the model makes online inference nearly three times faster. It also said that for agentic tasks, the overall cost of using the model is 70% lower than its predecessor.
MiniMax said it got the speed gains through "more aggressive sparsity." It said output speeds in everyday use can reach more than 200 tokens per second, with median speeds staying above 170 tokens per second.
The company listed flat prices per million tokens: $0.15 for input, $0.60 for output and $0.015 for cached input.
Why it matters: MiniMax is pitching the model on both performance and price. The company said M3.1 Flash performs on par with top models on three benchmarks while costing very little per task. The three are Automation-Bench, which tests complex agentic tasks, DeepSWE, which tests real-world coding, and Agents' Last Exam, which tests computer use on real-world tasks with economic value. MiniMax called the result "a highly competitive new standard for performance and cost."
The model is available now through the MiniMax API, a subscription called M Plan, and a desktop tool called MiniMax Code, the company said.
What is not known: The announcement does not give actual benchmark scores. It also does not say which rival models M3.1 Flash was compared against. None of the performance, speed or cost claims have been confirmed by anyone outside the company. The blog post does not name the predecessor model behind the 70% cost figure, and it does not say exactly what the "nearly three times faster" figure is measured against.
MiniMax did not say where Space Bunny was running while it operated anonymously.
The company said it will release the model weights later but gave no date. It also said a Pro version, which it described as pushing "the boundaries of intelligence," will come later, again without a timeline.
