Sber’s press service announced the first model of the upcoming generation of GigaChat, unveiled by Sber developers. When the training approach shifted, the model’s capabilities expanded across tasks such as condensing text, answering questions, and generating ideas. The system can now handle longer texts, effectively doubling the amount of content it can work with. The new model is accessible to all GigaChat users.
Officials reported that the user base has grown to about 1 million unique GigaChat users, and the launch marks a milestone for the new generation. The vice president for digital surfaces at Sberbank Salyut highlighted that the model not only mirrors human conversation but also shows marked improvements in productive tasks after additional training. In recent evaluations of neural network performance, notably the MMLU benchmark for multitask language understanding, GigaChat doubled its performance, signaling broad gains across capabilities.
In particular, the most notable progress occurred in summarization, an area that requires compressing large texts without sacrificing meaning. This enables users to extract essential insights quickly and efficiently. With summarized content, users can pose questions or shape important ideas, making it easier to find answers and articulate goals. Supporters expect that interactions with GigaChat will become more enjoyable and effective, with ongoing feedback driving further enhancements to the product.
Additionally, GigaChat can now be accessed on the VKontakte social network by activating the bot. The service leverages the Kandinsky model to generate images from text prompts and to produce creative ideas suitable for various situations. This expands the practical reach of the tool for daily tasks and creative work alike.
GigaChat is available to users across Sber smart devices and within the Salyut mobile app on Android. Businesses can also tap into the GigaChat API to build custom solutions and optimize processes, making it a versatile option for organizations seeking to harness advanced AI capabilities.