MiniMax H3 license: no training other AI models on its outputs

Under the MiniMax H3 Community License you may not use the Works or any of their Outputs "to improve any other artificial intelligence model (other than MiniMax H3 or its Model Derivatives)". That is Section V.3. If you keep generated clips as a dataset, that clause is the one to read before the dataset feeds a training run.

This quotes the license, read 2026-10-01; it is not legal advice, and Sume's docs do not address the clause. Sume facts below are from the Video generation docs.

How does the license define Outputs and Model Derivatives?

"Output" is any result of operating MiniMax H3 or its derivatives, including through Hosted Services. "Model Derivatives" include a model created by transferring the patterns of the weights or Outputs to another model so it behaves like MiniMax H3, including distillation and training on synthetic-data Outputs. The definition adds that Outputs themselves are not Model Derivatives.

Does VI.4 cancel V.3?

They sit side by side in the text. VI.4 says MiniMax claims no rights over Outputs, and V.3 restricts one use of them. How the two read together is for MiniMax and your counsel, not this post.

What does Sume say about stored clips?

Nothing about this clause. The docs describe the model ids (minimax-h3, 5 to 15 seconds at native 480p or 768p; minimax-h3-max) and job mechanics. Billing is reserved on submit at provider list times 1.25, and an Idempotency-Key makes a replay return the original job instead of creating a new one.

What should a team that trains models do?

Keep a record of which clips came from MiniMax H3 and exclude them from any training set unless MiniMax says otherwise in writing. Tag the model id in your own metadata when you store each job's result; see the async job pattern for how jobs return.