Sber open-sources 2B-parameter Kandinsky 5.0 video model
Sber has released Kandinsky 5.0 Video Lite, a 2-billion-parameter open-source text-to-video model built for efficiency, which it says outperforms earlier Kandinsky versions and most current open models.
Sber's Kandinsky team announced Kandinsky 5.0 Video Lite, the first model in its new Kandinsky 5 line. The open-source model has 2 billion parameters and generates video at 768×512. The team says it beats previous Kandinsky releases and most current open state-of-the-art models. Those quality claims, including comparisons with Wan 2.x on VBench and in side-by-side tests, come from Sber's own evaluations.
The model is a latent diffusion model trained with flow matching. Its backbone is a diffusion transformer that Sber designed in-house. Qwen2.5-VL encodes the text prompts, a CLIP embedding adds a global description of the prompt, and HunyuanVideo's 3D VAE compresses the video into latents. The core building block, called CrossDiT, uses cross-attention to text embeddings. It replaces the MMDiT design from Kandinsky 4.0, which relied on concatenation that the team says slowed training considerably. The team says the new design works better with sparse attention when videos of different lengths share a batch.
Sber also developed NABLA, a block-sparse attention method. The company says NABLA cuts training and inference time by up to three times without reducing quality, and it uses NABLA to generate 10-second clips.
The pretraining data started as 6 billion images and 35 million videos, which were split into 1.5 billion scenes of 2 to 60 seconds. After filtering, 124 million scenes and 520 million images remained. InternVL2, Qwen2.5-VL and Tarsier2 generated the captions. The team also hand-collected 230,000 scenes and 769,000 images of Russian cultural content. Pretraining runs in several stages, beginning with images at 256p before adding short videos.