Sber's Kandinsky Video ships open weights under Apache 2.0

The shot

Researchers from Sber's ai-forever group published the FusionFrames paper on 22 November 2023, describing Kandinsky Video, a text-to-video model built on the group's earlier Kandinsky 3.0 text-to-image system. The paper describes "a new two-stage latent diffusion text-to-video generation architecture," with a first stage generating keyframes to establish a video's storyline and a second stage generating interpolation frames to smooth motion between them. The project's own repository confirms this architecture and states the group followed it with an upgraded Kandinsky Video 1.1 that further splits keyframe generation into separate first-frame and subsequent-keyframe steps.

What the documents show

The paper reports that its pipeline achieves "top-2 scores overall and top-1 among open-source solutions" on stated metrics including CLIPSIM and Fréchet Video Distance, a comparison Sber's own team ran and reported. The repository adds a fact the paper does not need to state on its own terms: the code is released under the Apache License, Version 2.0, a fully permissive open-source license without a registration step, revenue threshold, or field-of-use restriction, and the corresponding Hugging Face model card carries the matching "apache-2.0" license field. That combination, an open architecture paper plus an unrestricted code and weights license, distinguishes Kandinsky Video from open releases elsewhere in this pack that pair open weights with usage-based commercial terms.

The workflow

Because the license imposes no revenue gate, a production or tool vendor can redistribute, modify, and deploy Kandinsky Video commercially subject only to ordinary Apache 2.0 attribution requirements, without the registration step some peer open releases require. The repository's own architecture details, including an 8.6-billion-parameter Flan-UL2 text encoder and a 4-billion-parameter latent diffusion U-Net3D, tell an integrator the model's compute footprint before deployment. The human decision that remains is quality verification against the paper's self-reported comparison, since Sber's stated ranking is not the same as an outside party independently reproducing it.

What the tool does not change

An unrestricted license changes what a team is legally permitted to build; it does not change what the model was shown to do in Sber's own reported evaluation, which remains the developer's self-benchmarked result until an independent test confirms or contradicts it. A permissive license and a strong self-reported score are two separate facts, and this record keeps them distinct rather than treating one as evidence of the other.

• Does a planned use rely on Kandinsky Video 1.0's original architecture or the later 1.1 update's added first-frame conditioning step?
• Has any party outside Sber's ai-forever group independently reproduced the paper's stated CLIPSIM and Fréchet Video Distance scores?
• What compute resources does the stated 8.6-billion-parameter text encoder and 4-billion-parameter diffusion model actually require for the deployment being planned?

Kandinsky Video shows that an open-weights release can pair a fully permissive license with a self-reported benchmark claim, and the license text settles what a team may build while leaving the performance claim exactly where the paper left it: unverified outside Sber's own report.