Sber unveils open-source AI suite for Russian language & media

Sber has released the weights of several new Russian-language neural network models, making them available under an open-source licence. The release includes flagship models from the GigaChat line, the latest GigaAM-v3 models for speech recognition, advanced models for image and video generation in the Kandinsky 5.0 range, and a new set of visual data compression tools, K-VAE 1.0.

AI model family

The new GigaChat Ultra Preview and GigaChat Lightning models are designed to cater to a broad range of tasks in Russian language processing. GigaChat Ultra Preview is described as the largest and most capable model in the range, while training is still ongoing. The company states that even at the current stage of development, it surpasses previous Russian-language AI models in overall quality benchmarks. GigaChat Ultra Preview is also being released under an open-source licence, allowing deployment in secure environments.

In contrast, GigaChat Lightning is the smallest and fastest model in the range, intended for local deployment and faster iteration cycles. Both models integrate systems for executing code and managing memory, enabling applications from automated calculation and data analysis to personalised advice that adapts as conversations evolve.

Open access

All new models, including those for image and video generation as well as speech recognition, are available under the MIT licence. This provides commercial and private organisations within Russia the opportunity to deploy, fine-tune, and build solutions on top of these models.

"Creating a true world-class AI requires two things: colossal resources and, even more importantly, world-class R&D teams. Sber has both. But our fundamental position is not to build a 'closed' technology. Our strategy is to become an open foundation for the entire country. This is precisely why we are open-sourcing our model weights. This is a key point. When we provide these models, any company in Russia-from a bank to a startup-can deploy them within their own secure perimeter and fine-tune them on their sensitive data without exposing it to anyone. This is what genuine technological sovereignty looks like: when the entire nation has access to AI, and it becomes the basis for business transformation and a stimulus for economic growth. I would also like to note that the Ultra model will soon be available to corporate clients as well, optimized for total cost of ownership for on-premises deployment within a company's perimeter," said Andrey Belevtsev, Senior Vice President, Head of Technology Development, Sber.

Speech recognition

The GigaAM-v3 suite consists of five automatic speech recognition models designed for Russian. These models were trained on 700,000 hours of audio data, a significant increase from previous versions. The training data covers a range of domains including call centres, musical queries, and spontaneous or accented speech. Sber indicates these models support a variety of business and industrial applications, from contact centre analytics to large-scale voice assistants.

Image and video generation

The Kandinsky 5.0 line introduces new capabilities in generative AI for image and video content, with models natively understanding prompts in Russian and generating Cyrillic text. Image Lite offers text-to-image generation and editing at HD resolution, supporting both Russian and English prompts. The suite's video models, Video Lite and Video Pro, support the creation of videos from text descriptions up to 10 seconds at 24 frames per second. These models were trained on nearly a billion images and 300 million videos, including content curated for relevance to Russian culture.

Visual data compression

K-VAE 1.0 is a set of autoencoders trained from scratch for compressing image and video data into latent spaces. According to Sber, these models support more efficient training and application of generative AI tools. K-VAE 1.0 is described as the best performing suite among open-source alternatives for this purpose and is now available for modification and use without restriction.