Yandex launches next-generation language models YandexGPT 4: Smarter reasoning and four times more text capacity

Yandex introduces the next generation of its generative neural networks: YandexGPT 4. This new lineup features a robust Pro version and a lighter Lite option. Both models outperform previous versions in response quality, reasoning capabilities, and processing power, handling up to four times more text — around 60 pages. With these upgrades, YandexGPT 4 can take on a wider range of business challenges, from analyzing customer inquiries to streamlining procurement processes.  

The new models are now available through the Yandex Cloud API, with some features still in beta. Businesses can use them to sort emails and client requests, analyze resumes, and manage other text-related tasks. The Pro version is perfect for complex tasks like sales analysis, while the Lite version is better suited for simpler scenarios where speed is a priority. You can test the new models in demo mode in the Yandex Cloud chat. Soon, they’ll be available in other Yandex services for a broader audience, starting with the Alice Pro option. 

Better responses

YandexGPT 4 Pro consistently outperforms the previous generation, delivering better results in 70% of cases on average. Its performance is nearly on par with GPT-4o in tasks like open-ended questions. The Lite version, thanks to improved training methods, matches the performance of Yandex’s most advanced model from the previous generation. In particular, the new models have been trained to reason through step-by-step problem-solving examples. 

Smarter reasoning

YandexGPT 4 introduces advanced chain-of-thought reasoning. Before answering complex questions, the new models break them down into smaller tasks and solve them step by step, building a chain of reasoning. This structured approach improves accuracy and allows the models to handle more analytical tasks. For example, the models can review a customer complaint, pinpoint the issue, and suggest a solution. In the future, the API will include an option to enable hidden reasoning for all queries.

Process more text faster 

The new models can process four times more text than the previous generation. As a result, they can maintain conversation context for longer, answer complex, lengthy questions, and analyze up to 60 pages of text. The new models provide better answers by drawing on external sources used in Retrieval-Augmented Generation (RAG) systems. These sources can include company documents or articles. The error rate, including hallucinated content, has been reduced by nearly half, from 4% to 2.1%. On top of that, the models now respond twice as fast as before.

Integration with third-party apps

YandexGPT 4 can generate commands for third-party apps. For example, if you ask it to find a plane ticket, it will generate the necessary command for a booking service to search for available seats. Command generation will soon be available in Yandex Cloud, allowing developers to use the models to work with third-party apps. Developers will only need to define the functions and rules for command creation, and the models will determine when to apply them on their own.

Contacts

Yandex Press Office
pr@yandex-team.com