MiniMax's free M3.1 Flash ties Claude Opus 5

MiniMax made its M3.1 Flash AI model free to download on the 23rd of October 2026, the Chinese company posted on X. It ties Anthropic's Claude Opus 5 on one of five tests, its model page shows.

The model has about 428 billion parameters and uses 23 billion of them for each token, MiniMax wrote. Parameters are the numbers a model learns, and a token is a piece of a word.

MiniMax's M3.1 Flash reads up to 1 million tokens of text, images and video

The model reads text, pictures and video, up to 1 million tokens at once, its Hugging Face page says. Hugging Face is a website that stores AI models for download.

The page calls it a mixture of experts, meaning it switches on only some parts for each token.

MiniMax's M3.1 Flash ties Anthropic's Claude Opus 5 on 1 test

M3.1 Flash scored 76.2 on BankerToolBench, the same as Anthropic's Claude Opus 5, the model page shows. On DeepSWE, a coding test, it scored 73.7 against 74.0.

On OSWorld, a test of using a computer, it scored 61.7 against 70.6, the page shows. On Terminal Bench 4.0, a test of work in a command window, it scored 38.1 against 51.8.

Below we break down the 428 billion and the scores

Using 23 of 428 billion parameters puts about 5.4% of the model to work for each token, by our arithmetic.

M3.1 Flash trails Claude Opus 5 by 13.7 points on Terminal Bench 4.0 and 8.9 on OSWorld, by our arithmetic. It leads by 12.5 points on Automation-Bench, a test of office tasks.

What is confirmed and what is not

The scores come from MiniMax's own page, and no outside test was checked. The licence, meaning the rules for using the model, was not checked.

Key facts

MiniMax released M3.1 Flash on the 23rd of October 2026.

It has about 428 billion parameters and uses 23 billion per token.

It reads up to 1 million tokens of text, images and video.

Anyone can download it from Hugging Face.