Open video model licences, Oct 2026: Kandinsky, Prism, H3, LTX

Of the open video models in the news this week, Kandinsky 6.0 Video and Tencent's Prism are MIT-licensed, while MiniMax H3 and its FastH3 Trim derivative use the MiniMax H3 Community License with territory and revenue conditions, and LTX-2.5 uses Lightricks' community license with a $10 million revenue line. "Open weights" does not mean "no conditions", so match the license to where you sell and how big you are before you pick a model.

What does each license say?

The table summarizes the pages I read on 2026-10-11. It is a reading aid, not legal advice, and each repository's own text controls.

Why does Prism need a second look?

Prism's license file is MIT, but it says third-party components stay under their original licenses; the one named example is Apache-2.0. Prism is also a research preview: the README lists preview alpha and beta checkpoints, and says 720p inference runs on one 80 GB NVIDIA GPU while 1080p and 2K need four or more. Native 2K is a lab setup, not a laptop job. Prism supports text-to-video and image-to-video, and the README lists 289 frames at 720p or 1080p and 265 frames at 2K, both at 24 fps. Training is far heavier than inference: the README asks for at least 32 GPUs of 80 GB at 720p and 64 at higher resolutions, so fine-tuning it is a cluster project.

Which license is the easiest to ship under?

For a small studio anywhere, MIT is the least friction. For a team in the US, EU, UK or South Korea, the H3 license text I read excludes those territories, so H3 and FastH3 Trim need a conversation with MiniMax first. For a company above $10 million, LTX-2.5 needs a paid agreement; above $20 million, H3 needs authorization.

Where does a hosted workflow fit?

A license decides whether you may run the weights. It does not give you a GPU, a queue or an edit step. Sume does not list Kandinsky or Prism in its repository; its Video Router docs list rows such as wan-3.0 (2 to 30 seconds), minimax-h3 (5 to 15 seconds at 480p or 768p) and gemini-omni-flash-1.1 (3 to 10 seconds). Use those when you want a clip with no hardware and a price quoted per job, and keep the MIT models for cases where you need the weights themselves.