GigaChat Description
GigaChat can answer questions, maintain a dialog, write code, and create texts and images based on descriptions in a single context. Unlike a foreign neuron network, GigaChat initially supports multimodal interactions and communicates better in Russian.
The architecture of GigaChat is based on a neural network ensemble called NeONKA (NEural Omnimodal Network with Knowledge-Awareness), which includes several neural network models, and the method of supervised tuning, reinforcement learning with feedback from humans. Sber's neural network can perform a variety of intellectual tasks, including maintaining a conversation, writing texts, and answering factual questions. The Kandinsky model 2.1 is included in the ensemble, giving the neural network the ability to create images.