Translated with AI from the Russian original.
Sberbank has presented the updated GigaChat 2.0 model line. According to the MERA benchmark, the flagship model GigaChat 2 MAX ranks first among AI models for the Russian language. In international tests, the new line outperforms GPT-4o, DeepSeek-V3, LLaMA 70B, and Qwen2.5 in a number of indicators.
The updated GigaChat 2.0 line allows businesses to solve a wide range of tasks faster and more efficiently:
· GigaChat 2 MAX -model with improved text processing, programming, and data analysis capabilities.Optimal for complex tasks and high loads.
· GigaChat 2 Pro -model with performance at the level of the previous MAX, but withlower requirements for computing resources.Distinguished by accuracy and creativity.
· GigaChat 2 Lite -basic version, comparable to the previous GigaChat Pro. Suitable for solving everyday tasks with reduced load on the infrastructure.
All models are available through the cloud API or can be deployed locally (on-premise)in the company's infrastructure.
GigaChat 2.0: key improvements
1. Improvement of AI agents
GigaChat 2.0 models are now better at:
· Autonomous solutions for complex multi-component tasks.
· Data analysis and creation of legal documentation.
· Code generation (Python and JS support, LangChain compatibility).
2. Four times more context
Now the neural network can analyze up to 200 pages of A4 in one request (previously - 48 pages). This makes GigaChat 2.0 an ideal tool for chatbots and complex analytical queries.
3. Performs work tasks 2 times better
· Follow user instructions twice as accurately.
· Answers questions 25% more accurately.
· Maintains style, formats answers to client's request.
According to the MERA benchmark, GigaChat 2 MAX ranks first among Russian-language AI models. In MMLU, HumanEval, and other international metrics, it is:
· A leader in factual questions in Russian.
· Best in exact sciences and logical thinking.
· Outperforms analogues in code generation and data structuring.
