Kandinsky 6.0 Video is MIT: can you use the clips commercially?

Yes on the headline question: Kandinsky 6.0 Video is published with its code and model checkpoints under the MIT license, per its paper and repository, and MIT has no revenue cap, territory list or attribution banner of the kind some other open video licenses carry. Read the notice file and any third-party component licenses in the repo before you ship, because the MIT text covers the project, not every dependency it pulls in.

What exactly is released?

The paper is dated October 6, 2026. It describes two models that generate picture and sound together from text or from an image: Kandinsky 6.0 Video Lite at 3 billion parameters and Pro at 29 billion. Clips are 5 seconds long. The base video is generated at SD resolution and upscaled to Full HD (1920 by 1080) by a separate super-resolution model, so "Full HD" is an upscale step, not native output.

The repository lists integrations with Diffusers, ComfyUI, vLLM-Omni, SGLang and FastVideo.

What does MIT give you, and what does it not?

MIT permits use, copying, modification and distribution of the software, including commercially, provided the copyright and permission notice travel with copies. It carries no warranty. It does not say anything about whether a particular clip infringes someone's likeness or trademark, and it does not replace platform rules or disclosure laws for AI-generated media in your market.

I did not audit the licenses of the text encoders or other components the install pulls in. Check them the way you would for any open-weights stack.

What does it cost to run yourself?

The repository gives one concrete figure: a Pro HD video takes about 292 seconds on an H100 and about 2,716 seconds on an RTX 5080. Setup needs Python 3.13 or 3.14 and installs a PyTorch build matched to your GPU. The paper mentions consumer-GPU deployment with block offloading and VRAM presets but the pages I read give no exact VRAM floor, so measure on your card before planning a batch.

Is it on Sume, and what is the practical choice?

I searched the Sume repository and found no Kandinsky model, so do not send it as a model id. Sume's Video Router docs list other rows: wan-3.0 accepts 2 to 30 seconds, minimax-h3 accepts 5 to 15 seconds at 480p or 768p, and gemini-omni-flash-1.1 is 3 to 10 seconds with native synced audio always on.

If you need exactly Kandinsky's output or want to fine-tune, run the MIT weights yourself. If you need a 7-second or 30-second clip, or no GPU to babysit, pick a hosted row and read its capabilities from the catalog first.