Sber launched ChatGPT competitor - GigaChat service

Sber has released its own version of a multimodal neural network that can answer questions

users, maintain a dialogue, writeprogram code, create texts and pictures based on descriptions within a single context, Rossiyskaya Gazeta reports. Unlike ChatGPT, the GigaChat service initially already supports multimodal interaction and communicates more competently in Russian. Access to the system is by invitation only. According to Sberbank, 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 the method of supervised fine-tuning, reinforcement learning with human feedback. Thanks to this, Sber's new neural network can solve many intellectual tasks: keep up a conversation, write texts, answer factual questions. And the inclusion of the Kandinsky 2.1 model in the ensemble gives the neural network the skill of creating images. ru-GPT 3.5, which is part of NeONKA, with 13 billion parameters, is planned to be made publicly available in order to accelerate the development of the open-source community and technologies in the field of artificial intelligence.

“GigaChat is unique in its open architecture,while the world analogues of our model adhere to the Closed AI course. It is important that GigaChat can be used not only by mass users who like to experiment with innovations, but also by students and even researchers for serious scientific work,” said German Gref, head of Sberbank. As explained in Sberbank, the Salyut virtual assistant is one of the first products that will receive all the features of the new model in the near future. The first version of the NeONKA model was optimized in size and speed for use in high-load services. The neural network was trained on the Christofari Neo supercomputer. The model is currently in closed beta mode.