Can you train a model on open video weights' outputs? Four licenses

H3 and HunyuanVideo bar using outputs to improve other models. LTX bars commercial distillation. Wan 2.2 is Apache 2.0. The clauses, quoted from vendor files.

5 min readSume
All posts

Not freely. MiniMax H3's FAQ says outputs and weights may not be used to train or improve other AI models, HunyuanVideo's license says the same in Section 5(b), and the LTX-2.x license says training, fine-tuning or distilling a model for commercial use is not a Non-Commercial Purpose. Only Wan 2.2, which its README places under Apache 2.0 and claims no rights over your generated content, carries no such clause in the files read. All four read from vendor pages on 2026-10-08. This is a reading aid, not legal advice.

The clause by model

Each cell is the vendor's own wording, shortened.

Training on outputs: vendor wording, read 2026-10-08
ModelWhat the vendor text saysWhere
MiniMax H3Outputs and weights may not be used to train or improve other AI models; distillation into non-H3 models is not permittedH3 FAQ, commercial-use answer
HunyuanVideoYou must not use the works or any Output or results to improve any other AI model other than Hunyuan or its Model DerivativesLicense Section 5(b)
LTX-2.5Training, fine-tuning or distilling a model for commercial use is not a Non-Commercial Purpose and needs the paid agreement for entities at $10,000,000 revenueSection 2.2 of the August 11, 2026 license
Wan 2.2Apache 2.0; the README says Wan claims no rights over your generated contentREADME license section

Two traps in the wording

First, Hunyuan's definition of Model Derivatives includes models created by transfer of patterns from the weights or Output, including methods based on generating synthetic data. It also states that Outputs by themselves are not Model Derivatives. So the output file is not a derivative, but using it as training data for another model is still barred by Section 5(b).

Second, H3's FAQ separates two things. LoRAs and fine-tunes of H3 are allowed and owned by you as derivative models. Training a different model on H3 outputs is the part that is barred. A style LoRA for H3 is fine; a classifier or a rival video model fed with H3 clips is not.

Hosted services and the same question

A hosted API has its own terms, separate from the weights license. Sume's video docs describe the job flow, pricing and the lack of a zero-data-retention toggle; they do not state a rule about training other models on outputs, so this post makes no claim either way. Read the terms of the service before you build a dataset from hosted output, and read the underlying model vendor's terms too.

A working rule for dataset teams

Before you add generated clips to a training set:

  • Write down which model made each clip and under which license or service terms.
  • Exclude H3 and Hunyuan outputs from datasets for any non-derivative model.
  • For LTX, check your entity's revenue and whether the use is commercial.
  • For Wan 2.2, read LICENSE.txt and the README's list of prohibited uses before relying on 'no rights claimed'.

Sources

Related posts

More in Models

All Models posts

Written by Sume