Kandinsky S2V 1.0: Open-Weight Talking Avatar Models Released Under MIT License

Kandinsky Lab, the team behind Sber’s Kandinsky models, has released Kandinsky S2V 1.0, a family of “speech-to-video” models that transforms an image and a voice track into a video of a talking person. The code is available on GitHub and the weights on Hugging Face; the specialized outlet mlllm.io reported on it on November 3, 2026. Both the code and the weights are under the MIT license, allowing for commercial use. That is rare for this type of tool, which is currently dominated by proprietary avatar and lip-sync services.

The whole body, not just the mouth

The README emphasizes one key point: the entire figure is animated, not just the head. Using a reference image, a voice recording, and a short description, the model generates video at 25 frames per second where lip movements, facial expressions, posture, and gestures align with the speech. Output is available in standard definition or HD, matching the aspect ratio of the initial image. Extended speeches are generated in 5-second blocks connected by a memory buffer of previous frames.

Two sizes are offered: Lite with 2 billion parameters and Pro with 21 billion, each featuring a faster distilled version. Technically, the models build upon Kandinsky 5.0 Video and use Qwen2.5-Omni to process both text and audio, replacing standard audio encoders. The team reports training them on 1.17 million single-speaker clips, captioned in both English and Russian. Integration with the Diffusers library exists, though it has not yet been rolled into a public release.

Open, but demanding

While the license is open, the hardware requirements are decidedly steep. The repository calls for an NVIDIA GPU with 48 GB or more of VRAM, and 80 GB for HD. A decoder spatial tiling option lowers the requirement to around 28 GB, though with slightly visible seams. Even distilled, the Lite version takes nearly 9 seconds to produce a single second of video on an H100. There is no API and no pricing: this is a research release, accompanied by an online demo on Hugging Face.

Regarding output quality, the figures come strictly from the team. They claim that in blind comparisons, the Pro version performs on par with or better than the leading open-source model as well as two unnamed commercial offerings, and that it is favored for lip sync. No independent benchmarks have been published yet.