Sberbank has introduced Kandinsky 6.0 Video, a multimodal model capable of generating videos with synchronized audio, including speech, music, and ambient sounds. The model is integrated into GigaChat, allowing users to create content in Full HD resolution for free.
The system generates audio at 44 kHz, matching standard streaming quality. Users can create clips up to five seconds long by providing text descriptions or an initial frame. The model features improved physics simulation for more natural movement of people, animals, and objects.
Kandinsky 6.0 Video is available to developers under the MIT open-source license. It can be integrated into existing products via platforms such as FAL, Diffusers, FastVideo, ComfyUI, SGLang, and vLLM-omni, without requiring custom integration from scratch.
The model's audio module was trained on over 20 million video clips. According to internal benchmarks, Kandinsky 6.0 Video outperforms its predecessor, Kandinsky 5.0, in 71% of cases and surpasses the LTX 2.5 model and the Veo 3.1 Fast model in specific visual and motion tasks.
