What actually launched on October 6
Sber's Kandinsky Lab released Kandinsky 6.0 Video, a pair of diffusion models that generate a video clip and its soundtrack at the same time, including dialogue, music, ambient sound, and lip movement that matches the audio. The lineup includes a 29-billion-parameter Pro model and a lightweight 3-billion-parameter Lite model, and both are released under the MIT license with code, weights, and Diffusers integration available for free, commercial use.
That licensing choice is the real headline. Most 2026 video model releases have stayed closed or API-gated, so a model that generates synced audio and lip-sync in a single pass, and that you can self-host or build a product on top of without restriction, stands out as unusual for this moment in the market.
How it works and what it costs
You describe a scene in text, or upload a first frame along with an audio description, and the model generates a clip with sound synced to the frame and lip movement matched to dialogue. The catch is clip length: the maximum duration with synchronized sound is five seconds. The developers have said they plan to extend this over time, but as of this release, five seconds is the ceiling.
Access is genuinely low-friction. There's a free Hugging Face demo, ComfyUI nodes, and free generation inside Sber's GigaChat app with SD, HD, and Full HD output options. A plug-in super-resolution model can lift output to 1920x1080, since the base clip renders at lower resolution first. On the hardware side, the int8 build runs synchronized video-with-audio generation on as little as 8 GB of VRAM, which puts it within reach of a lot of consumer GPUs, not just data-center setups.
How it stacks up against existing models
In side-by-side human evaluation, Kandinsky 6.0 Video Pro clearly outperforms its predecessor, Kandinsky 5.0 Video Pro, and Sber says the new model beat version 5 in a majority of comparisons. The company also cites improved real-world physics handling, meaning movement of people, animals, and objects should read as more natural than before.
Against outside competition, human evaluators preferred Kandinsky 6.0 Pro over LTX 2.5 on visual quality, motion realism, prompt following, artifact reduction, and overall task performance, with a statistically significant edge in speech quality specifically. The brief notes it remains competitive with leading audio-video generation models generally, particularly on speech. That's a meaningful claim for a free, open-weight model to make against paid competitors, though it's worth remembering these are the lab's own cited evaluations rather than independent third-party benchmarks.
Where this fits in a short-form workflow
Five seconds of synced audio and video is not enough for a full episode, a sketch, or even most single-scene stories. What it is good for is the stuff that normally eats time without being the point of the video: B-roll with ambient sound already baked in, reaction inserts with a lip-synced quip, or short cutaways you'd otherwise build by stitching together a video generator, a separate voice tool, and manual sound design.
That one-pass workflow is the actual value here. Normally, getting a lip-synced talking clip with matching sound effects or music means at least three tools and several rounds of alignment. Kandinsky 6.0 Video collapses that into one generation, even if the output is short enough that you'll need to chain multiple clips together for anything longer.
Because it's MIT-licensed and free to run, it's also a reasonable place to test character voice-and-lip-sync fidelity before spending credits on paid platforms like HeyGen or Kling. If a character concept or line reading doesn't work in a free five-second Kandinsky test, you've learned that without burning paid generation credits.
- Good fits: B-roll with built-in ambient sound, reaction inserts, short lip-synced cutaways, early voice and lip-sync testing
- Not a fit yet: full scenes, multi-line dialogue exchanges, anything needing more than five seconds of continuous synced audio
- Access points: free Hugging Face demo, ComfyUI nodes, GigaChat app (SD, HD, Full HD), runs locally on 8 GB VRAM with the int8 build
What's coming next, and what's still unclear
The engineering team plans to roll out compatibility with FAL, Diffusers, FastVideo, ComfyUI, SGLang, and vLLM, which would let developers integrate synchronized audio generation directly into custom video pipelines without commercial restrictions. If that lands, it widens the gap between Kandinsky and closed, API-only competitors for anyone building tooling rather than just generating one-off clips.
What's not yet clear from available reporting is a firm timeline for extending past the five-second limit, and how the model performs on more complex multi-speaker dialogue or non-Russian language lip-sync at scale. The brief's sources are consistent on the current five-second cap and the evaluation results cited above, but longer-term roadmap details are thinner, so treat any extended-length claims as unconfirmed until the lab publishes more.
Should you test it now
For short-form creators, the case for trying Kandinsky 6.0 Video is straightforward: it's free, it runs on modest hardware or entirely in the browser via Hugging Face or GigaChat, and it does something no other fully open, commercially unrestricted model currently does in one pass. The five-second limit means it won't replace your main video pipeline, but as a supplement for synced-audio cutaways or quick lip-sync experiments, there's no real cost to testing it before you commit credits elsewhere.
If you're organizing a workflow around multiple short AI-generated clips, a tool like Episodo can help you plan and sequence those pieces into a finished short, which matters more now that five-second synced clips are realistically going to be stitched together rather than used alone.