Kandinsky 5.0

Kandinsky 5.0 covers the whole local video stack in one family: Video Lite 2B checkpoints for 5 and 10 second clips at 768x512, Video Pro at 19B pushing quality up to 1920x1080, plus a 6B Image Lite text-to-image model and a 6B editing variant, all latent diffusion transformers trained with flow matching. Kandinsky Lab published unusually complete assets, including pretrain checkpoints alongside SFT, CFG-free, and distilled versions of the Lite models, so researchers can build on intermediate stages rather than only polished final weights. Speed spans the range too, from 13-second image generation to around 35 seconds for a distilled Lite clip and about 20 minutes for a 5 second 1080p Pro clip on an H100. Everything is MIT licensed with code on GitHub, weights on Hugging Face, Diffusers support, and ComfyUI integration; offloading and quantization options stretch requirements down to roughly 12GB VRAM for the smaller models. The staged release ran September through December 2025 and remains among the most open video model programs anywhere.