Kandinsky Lab · Open Audio-Video Model

Kandinsky 6.0 Video

An open-source family of diffusion models, released 6 October, that generate 5-second clips with synchronized 44 kHz audio, including lip-sync, from text or an image plus text. It comes in Lite (3B) and Pro (29B). A dual-stream CrossDiT joins the Kandinsky 5.0 video stream to a newly trained audio stream through bidirectional cross-attention, and a plug-in super-resolution model lifts output to 1920x1080; Pro defaults to 864x480 at 24fps for 125 frames. MIT, with pretrained, full and distilled checkpoints, a Hugging Face Space, and day-one support in Diffusers, ComfyUI, vLLM-Omni, SGLang and FastVideo. Kandinsky Lab reports Pro beating 5.0 Pro in human evaluation, self-reported. Kandinsky Lab's homepage was covered in week 47 of 2025.