KandinskyLab develops a family of open-source generative AI models under the Kandinsky name, covering text-to-image, image editing, text-to-video, and image-to-video synthesis. The lab has released a series of progressively more capable model generations since 2021, with the latest Kandinsky 5.0 family built on Flow Matching architecture and scaling from 2 billion to 19 billion parameters.

The Kandinsky 5.0 lineup includes three model tiers: Image Lite, a 6-billion-parameter model that generates high-definition images at up to 1280x768 and 1024x1024 resolution with strong text rendering and Russian concept understanding, including an image editing variant; Video Lite, a lightweight 2-billion-parameter video model ranked first among open-source models in its class; and Video Pro, a 19-billion-parameter model delivering high-quality HD video at 1280x768 and 24 fps with rich motion dynamics and precise camera control.

All Kandinsky models support both English and Russian text prompts. Prior generations introduced capabilities including ControlNet-based editing, keyframe interpolation, cultural awareness for Russian-language prompts, and audio generation conditioned on video. The lab actively publishes its research through peer-reviewed venues including EMNLP, NAACL, and IEEE Access, and releases model weights and code openly on GitHub and Hugging Face.