Sber open-sources full Kandinsky 5.0 lineup
Sber has released the rest of its Kandinsky 5.0 image and video models as open source, including a 19B video model it says matches Veo 3 and is the strongest open model available.
Sber's Kandinsky team detailed the open-source release of the remaining Kandinsky 5.0 models in a technical post on Habr. The company first announced the release a week earlier at its AI Journey conference. It follows September's Video Lite, a 2B text- and image-to-video model. Since then, Sber has added support for consumer GPUs with 12 GB or more and code for LoRA fine-tuning.
The newly opened models are:
- Image Lite (6B): HD image generation and editing.
- Video Pro (19B): text-to-video and image-to-video, producing clips of up to 10 seconds at resolutions up to 1024×1024, 1408×640 or 1280×768. Video Lite tops out at 512×768.
Video Pro uses 60 CrossDiT diffusion-transformer blocks, against 32 in Lite, and widens the embedding dimension from 1,792 to 4,096. Sber says its in-house NABLA sparse-attention mechanism helps make long, high-resolution generation workable. Pretraining used more than 250 million video scenes and mixed text-to-image, text-to-video and image-to-video tasks, with resolution raised from 256px to 1024px. Supervised fine-tuning used expert-curated data and "model souping," which merges the weights of models trained on separate domains. Distilled variants cut sampling from 100 steps to 16 using CFG distillation and Trajectory Segmented Consistency Distillation, followed by adversarial post-training.
Sber says Video Pro clearly outperforms Wan 2.2 A14B and matches Google's Veo 3 on visual quality and motion. Those comparisons are the company's own, and the post excerpt includes no benchmark figures or independent evaluation. All of the models accept Russian and English prompts and can render text in Cyrillic and Latin script. Licensing terms are not stated in the excerpt.