MiniMax M3.1 Flash

A new fast AI model

MiniMax has officially launched M3.1 Flash. It is a "Flash" model, which means it is built to be quick and efficient. The company says it makes big gains in agentic tasks (where the AI works on multi-step jobs by itself), coding, computer use, and cybersecurity. An early version, nicknamed Space Bunny, handled more than 200 trillion tokens while it ran without a public name.

Key numbers

• 73.7% on DeepSWE v1.1, a test of real software tasks, placing it alongside top models.
• 61.7 points on OSWorld V2, a test of AI operating computers across apps. It ranks first among Chinese models on this test.
• 29.6 points on Agents' Last Exam (ALE), a test of real professional work. It also ranks first among Chinese models here.
• 82% on CyberGym, a score for finding and checking security weaknesses in software.
• Speed: over 200 tokens per second at peak, and a typical (median) speed of 170+ tokens per second.

What tokens and benchmarks mean

A token is a small piece of text that an AI reads or writes. Companies measure AI models with benchmarks, which are standard tests that compare models. The article's results come from the company's own reports, so they are not independent checks. Higher scores on these tests usually mean the model handles harder problems better.

Why it matters

Cheaper and faster AI models can make it easier for businesses and developers to use AI for everyday work, such as writing code, handling customer tasks, or checking security. If the claims hold up, M3.1 Flash could offer top-level results at a lower cost. The company also says it will release the model weights later, which would let others study and run the model themselves.

What to watch

• When the model weights are released to the public.
• When the more powerful Pro model arrives, which the company says will push the limits of intelligence.
• Independent tests that check whether the company's scores hold up in real use.