MiniMax's M3.1-Flash Weights Come With a $20 Million Revenue Line and a Military-Use Ban

MiniMax published the weights of M3.1-Flash on Oct. 23, 2026 under a custom licence that allows commercial use only if the user displays "Built with MiniMax M3.1-Flash" and sends the company a notice. Products that earn more than $20 million a year need MiniMax's prior written authorisation.

The terms are in the LICENSE file on Hugging Face, where the licence tag reads minimax-community. The repository's model card lists about 428B total parameters with 23B active per token and a 1M-token context window. Vals AI lists a release date of Oct. 20, and AiCybr reports that MiniMax announced the hosted model on Oct. 19 and published weights four days after.

What the MiniMax licence requires

The licence grants use "for non-commercial purposes" by default. It defines commercial use as any use of the model or a derivative "primarily intended for commercial advantage", and that includes paid offerings and commercial API use.

Condition	Text in the licence
Under $20 million a year	Display "Built with MiniMax M3.1-Flash" on a related website and send a one-time notice to api@minimax.io
Over $20 million a year	Prior written authorisation from MiniMax
Fine-tuned models	Deploying post-trained or fine-tuned versions for commercial purposes counts as commercial use
Prohibited uses	Includes any use to "assist with, engage in or otherwise support any military purpose"

The file contains no termination clause. The weights are downloadable without a gate form.

How it compares with MIT and Apache releases

DeepSeek put V4.1-Flash-1017 on Hugging Face under the MIT licence on Oct. 17. The site's earlier coverage of Recursive's Flint covered an open-weight coding model under Apache 2.0. The MiniMax licence text adds an attribution requirement, a revenue threshold and a military-use ban to the same kind of download.

A startup under $20 million in revenue can ship a MiniMax derivative with a banner and an email. A company above that line has to ask first. The MIT licence has no such step.

MiniMax's benchmark claims against Vals AI

The card lists a Terminal-Bench 4.0 score of 38.1. Vals AI's leaderboard lists 25.25% for the same benchmark, ranked 26th of 57. The card's figures are vendor-supplied.

1. DeepSeek V4.1-Flash-101761.26 %
2. MiniMax M3.1-Flash53.78 %

Source: Vals AI leaderboard pages for each model, accessed 2026-10-29

Vals ranks MiniMax 19th of 29 on its Vals Index at 53.78%. It ranks 42nd of 121 on Vibe Code Bench v1.1 at 73.94%. Vals lists token prices of $0.15 per million input tokens and $0.60 per million output tokens. On Tax Agent Bench, a different test from Surge's Dayjob: Tax, it scores 32.14%, 21st of 78.

Vals shows two cost figures for the Vals Index on the same page, $0.660 per test in the specification section and $4.46 in an Oct. 20 update. The page does not reconcile them.

What it takes to run

The checkpoint is about 252 GB on Hugging Face, according to AiCybr, which cautioned that file size is not a measured memory requirement. The card names SGLang, vLLM and TokenSpeed as serving frameworks and lists no hardware minimum. The repository's file metadata shows 220B parameters, which does not match the card's 428B.

AiCybr also cites VentureBeat's Oct. 19 coverage of Artificial Analysis testing: an Intelligence Index score of 42 and a 93% rate on its AA-Omniscience hallucination measure. The VentureBeat and Artificial Analysis pages were not read directly, so those figures are as AiCybr reports them.

The licence file's text on the $20 million line reads "more than 20 million US dollars (or equivalent in other currencies)" in yearly revenue, so the threshold applies to products, not to the whole company's accounts. The card shows no gating notice, so the 252 GB download is available to anyone willing to accept the terms above.