MiniMax put M3.1 Flash out on October 19 with a price that makes the parameter count feel like a dare. Fifteen cents for a million input tokens is cheap enough that someone will try it this week for the wrong job.
The spec, without the poster
The model is a mixture of experts: 428 billion parameters in total, about 23 billion of them active when it answers. That split is why a model this large can be priced like a small one. Context goes to a million tokens. Artificial Analysis moved it from a 29 to a 42 on its Intelligence Index compared with the earlier M3.
MiniMax said the weights would follow on October 23. A weight release is not the same as a permission slip. Read the license on the day you download it, especially if the output is going into a product you sell.
Where the cheap token lies
A low input price rewards long prompts and long documents. It does not promise that the answer is the one you wanted. Use it first on work you can check: a refactor with tests, a summary of a paper you have open, a translation you can spot-read.
If the weights land on schedule, local and private deployments get interesting for teams that cannot send client text to a US lab. Interesting is not the same as ready. Run your own tasks before you migrate a workflow because a leaderboard twitched.