Kandinsky S2V 1.0 ships 2B and 21B open speech-to-video models

The Kandinsky Lab released Kandinsky S2V 1.0, an open family of speech-to-video models that turns one reference photo and a speech track into talking avatar video at 25 frames per second.

The team posted the technical report to arXiv on October 13, 2026, and dates it October 15. Both models take a reference image, a speech waveform, and an optional text prompt, then generate video in which lip motion, expression, pose, and hand gestures follow the speech, according to the paper. The project page shows clips up to 77 seconds long and says the models also drive cartoon characters and animals, not only photographed people.

The family ships in two sizes. Kandinsky S2V 1.0 Lite has 2 billion parameters and Kandinsky S2V 1.0 Pro has 21 billion. Both pull text and audio conditioning from a single Qwen2.5-Omni model and add a three-scale visual memory that keeps the reference image in view across chunks. Training used 1.17 million single-speaker clips captioned in English and Russian, and 25.8% of that corpus is Russian speech, so Russian prompts work without extra fine-tuning.

A distilled variant, pi-Flow, generates video in 16 sampling steps without classifier-free guidance and runs 5.4 times faster end to end than the base Pro model, while matching or beating it on nearly all reported metrics, the paper states.

The paper names the benchmarks and the rivals. On EMTD, CelebV-HQ, and HDTF, the distilled Pro model takes the best mean rank and min-max score across five metrics, ahead of LongCat 1.5. The human study ran side-by-side comparisons against LongCat 1.5 and two paid systems, HeyGen Avatar IV and Kling Avatar 2. Pro was preferred for lip sync in 54% to 62% of judgments, and its one clear loss was visual quality, to LongCat 1.5 at standard definition.

Those two commercial services sell avatar video by the second. HeyGen lists its Avatar IV photo avatar at $0.05 per second, and Kling lists its Avatar at $0.056 per second at 720p. Kandinsky S2V 1.0 has no per-second price to run yourself, but it is not instant: the paper measures a distilled Pro HD chunk at 145 seconds on eight H100 GPUs.

The paper says the code, checkpoints, and a Diffusers integration are released under the MIT license. As of October 15 the GitHub repository and Hugging Face collection named in the paper did not resolve, and the lab's Hugging Face account carried no Kandinsky S2V 1.0 checkpoints, only two tiny test pipelines for Diffusers tests.

The report is on arXiv, and the release is headed for the Diffusers library; watch the Kandinsky Lab's GitHub and Hugging Face pages for the full weights.