GigaChat2.0 introduced in Russia

Sber has introduced GigaChat 2.0, a functional neural network model in Russian.

The bank's press service said that GigaChat 2.0 allows companies to create more productive autonomous assistants (AI agents) capable of reasoning and solving complex tasks on their own. The models improved their knowledge in mathematics, natural sciences, and humanities, as well as improved their programming and writing of high-quality code. To develop agents in Python and JS, it is acceptable to use the LangChain SDK, with which GigaChat is fully compatible. Compatibility packages are available in the GigaChain public repository. The new generation models retain the context of a conversation longer and analyze more text (the maximum query volume has grown to almost 200 pages), as well as follow user instructions twice as accurately and solve work tasks and answer questions 25% better: they follow the specified formats and conditions, form answers in a certain style.

According to Andrey Belevtsev, Senior Vice President and Head of the Technological Development Unit at Sberbank, "GigaChat 2.0 is not just an increase in metrics and technical characteristics, but a significant step in the development of Russian-speaking large language models (LLM)." "We have created a model at the level of the world's best solutions, and in tasks in Russian the model surpasses most of them. Strong domestic neural networks are strategically important for any business operating in Russia. 15,000 external customers are already using GigaChat, and a powerful line update will allow even more customers to more efficiently solve a huge range of tasks," Belevtsev emphasized.

Users retain access to the first-generation models and can test GigaChat 2.0 before upgrading to a new model range.

According to the MERA benchmark for the Russian language, the GigaChat 2 MAX model ranks first among AI models, and in international benchmarks, the updated model range surpasses GPT4o, DeepSeek-V3, LLaMA 70B and Qwen2.5 in many respects.