Sberbank presents GigaChat: a new neural network for multilingual communication and creative tasks

Sberbank has unveiled its new neural network, called GigaChat. At the first stage, it will be available in testing mode by invitation. This neural network can be used to ask questions, maintain dialog, write program code, and create texts and pictures based on descriptions within a single context. At the same time, GigaChat supports multimodal interaction and is able to communicate more competently with users in Russian.

The GigaChat architecture is based on the neural network ensemble of the NeONKA (NEural Omnimodal Network with Knowledge-Awareness) model, which includes various neural network models and methods of supervised fine-tuning and reinforcement learning with human feedback. Thanks to this approach, the neural network is able to solve many different tasks. In addition, the service uses a model of Kandinsky 2.1which allows the neural network to create images.

Sber's new development, which we presented today, is a breakthrough for the entire large universe of Russian technologies. Moreover, GigaChat is unique in its open architecture, while global analogs of our model stick to the Closed AI course. Importantly, GigaChat can be used not only by mass users who like to experiment with innovation, but also by students and even researchers for serious scientific work.
Herman Gref, President, Chairman of the Management Board of Sberbank.

One of the first products that will receive all the benefits of GigaChat is the Salute virtual assistant. Sberbank claims that the new smart neural network will be useful for professionals from various fields, such as media representatives, editors, programmers, teachers, students, marketers and others.

In order to start testing GigaChat, you need to join the to a closed community on the Telegram channel.