Kandinsky Neural Network Tops International Ranking of Generative Video Models
The Kandinsky neural network, developed by Sberbank, has taken first place among all open-source solutions on the international LMArena platform. LMArena, created by researchers at the University of California, Berkeley, operates as an open tournament in which users test large AI models, compare their capabilities, and collectively determine the rankings based on real-world evaluations.
Competing With Global Flagships
International experts gave high marks to Kandinsky 5.0 Video. In its Kandinsky 5.0 Video Pro version, Sberbank’s development emerged as the top performer among open-source competitors. Because it follows open-source principles, specialists can freely study the code, modify it, and adapt the model to their own use cases.
Kandinsky is a service for generating images and videos that works with prompts in Russian and English and is available at no cost. Users only need to describe an idea to receive a finished visual result, with composition, narrative, and mood shaped by the prompt. The model can generate video clips up to 10 seconds long, supports camera motion control, and can create video either from text descriptions or from a source image. Its open code base also makes it suitable for commercial projects.
Experts on the LMArena platform ranked Kandinsky 5.0 Video Pro alongside flagship products from leading global developers. Only closed models from major Chinese and Western technology companies ranked higher, including solutions developed by Google DeepMind, OpenAI, Alibaba, and KlingAI. At the same time, the Russian model demonstrated parity with Ray 3 from Luma AI and Hailuo 2.3 from MiniMax.
Recognition From the Research Community
The scientific community also highlighted the strengths of Sberbank’s neural network. A detailed technical report on Kandinsky 5.0 was presented at the international AI Journey conference and quickly rose to the top of the Daily Papers rankings on the Hugging Face research platform, leading first by day, then by week, and ultimately by month.
As part of the AI Journey conference, Sberbank released all image and video generation models from the new Kandinsky 5.0 lineup into open access, including Video Pro, Video Lite, and Image Lite. These advanced models incorporate elements of Russian cultural context, allowing them to accurately interpret prompts in Russian and generate Cyrillic text within images and videos. Another notable step was the open release of the K-VAE 1.0 model for visual data compression.
Setting a Direction for the IT Industry
International recognition of Kandinsky on the LMArena platform plays an important role in raising the visibility of Russian IT products globally. It demonstrates the high level of expertise among Russian AI developers and their ability to compete at the forefront of international innovation.
Equally significant is that Sberbank’s digital solutions remain open for community use, combining high technical quality with an open-source philosophy. Recognition of Kandinsky provides momentum for further development of generative AI in Russia and, over time, may support the growth of an AI ecosystem and startups built on this technological foundation.
