On October 6, 2026, Sber and the Kandinsky Lab team released Kandinsky 6.0 Video, a model that generates video and sound together. Sber calls it Russia’s first neural network with synchronized audio — the previous Kandinsky 5.0 produced silent clips. The model is free to use in GigaChat, and the code and weights are open under the MIT license.
What Kandinsky 6.0 Video can do
- Video with audio. Sound isn’t layered onto a finished clip — the same model generates it alongside the picture: characters’ speech with lip-sync, music and ambient noise. The track is 44 kHz, and audio can be turned off.
- Two modes: text-to-video, or a first frame plus a description.
- Length: 5 seconds, 121 frames at 24 fps.
- Resolution: the model renders at 864×480, and a separate Kandinsky 6.0 Video SR model upscales to 1920×1080. GigaChat offers SD, HD and Full HD.
- Unfamiliar objects. The system pulls in references for objects and characters that appeared after the model was trained.
Lite vs Pro
The family has two models. Lite has 3B parameters and is fast. Pro has 29B parameters and is the quality flagship. Both support both modes, and each has a distilled version that generates faster.
In the developers’ blind side-by-side evaluation, Kandinsky 6.0 Video Pro clearly beats Kandinsky 5.0 Video Pro and holds its own against leading audio-video models, especially on speech quality. Sber says the new version wins over the old one in 71% of cases on average.
Where to run it
- GigaChat: free, in the browser and the app.
- Hugging Face and GitHub: weights for Lite, Pro, the distilled versions and the upscaler; the MIT license allows commercial use.
- Locally: through Diffusers, ComfyUI or vLLM-Omni. Pro runs on a GPU with 16 GB of VRAM.
Limitations
Length is the big one. Five seconds with no extension mode means a 30-second piece takes six generations plus stitching. For comparison, Veo 3.1 makes 8 seconds in native 1080p, Grok Imagine goes up to 15 seconds, and Seedance 2.5 does up to 30 seconds in one pass. Kandinsky’s Full HD is upscaled from 480p rather than rendered natively. The technical report compares the model with Seedance 2.0, even though Seedance 2.5 was already out at release.
The previous Kandinsky 5.0 ranked 37th and 44th out of 48 on the Arena video leaderboard — more in our overview of Russian AI video generators. Kandinsky 6.0 hasn’t appeared in independent rankings yet.
How to try it in AI2Face
Kandinsky 6.0 Video is available in the AI2Face generator next to models that make longer clips with audio in native 1080p: Veo 3.1, Kling, Seedance and Gemini Omni. You can run one prompt through several models and keep the best take, and the cost is shown before each generation. See all options in our guide to AI video generators.
FAQ
Where can I try Kandinsky 6.0 Video?
For free in GigaChat. The Lite and Pro weights are on Hugging Face and GitHub under the MIT license, and Pro runs locally on a GPU with 16 GB of VRAM.
How long are Kandinsky 6 videos?
5 seconds: 121 frames at 24 fps. There's no extension mode yet; Sber says longer clips will come in future versions.
Does Kandinsky 6.0 Video generate audio?
Yes. It creates sound together with the video: lip-synced speech, music and ambience at 44 kHz. Audio can be turned off.
Is Kandinsky 6 better than Veo or Seedance?
Not on length or resolution: Veo 3.1 makes 8-second clips in native 1080p, and Seedance 2.5 does up to 30 seconds in one pass. The developers benchmarked Kandinsky 6.0 against Seedance 2.0, not the current 2.5.
Sources
5
- https://kandinskylab.ai/
- https://github.com/kandinskylab/kandinsky-6
- https://huggingface.co/kandinskylab/Kandinsky-6.0-Pro-5s-Diffusers
- https://arxiv.org/abs/2610.05608
- https://www.vedomosti.ru/technology/news/2026/10/06/1234628-kandinsky-dlya-video
