Яндекс запускает YandexGPT 4: новое поколение языковых моделей

New models can reason and process four times more text

Yandex is launching a new generation of YandexGPT 4 language generative models, which include both a powerful Pro model and a lightweight Lite version. These models significantly outperform previous versions in the quality of responses, are capable of reasoning, and can process text up to 60 pages in length, making them ideal for solving complex business tasks such as analyzing customer requests and automating procurement.

The models are already available via API in Yandex Cloud, and although some features are still in test mode, they can be used in various business areas, including sorting emails and analyzing resumes. The Pro version is designed for labor-intensive tasks such as sales analysis, while Lite is optimized for simpler scenarios where speed is important. You can try out the new models in a demo version on the Yandex Cloud platform, and in the future they will appear in popular Yandex services, the first of which will be Alice with the "Pro" option.

YandexGPT 4 Pro demonstrates a 70% improvement in the quality of responses compared to the previous model, especially in tasks with open-ended questions. Thanks to a new training process, the models have learned to reason, breaking down complex tasks into simpler steps and building a chain of reasoning, which allows them to offer informed solutions, for example, in the case of analyzing customer complaints.

The new models are capable of processing four times more text than their predecessors, which significantly improves context retention and allows them to respond to long and complex queries. The reduction in the proportion of errors in responses is also almost two times, from 4% to 2.1%. YandexGPT 4 responds to requests on average twice as fast as previous versions.

In addition, the new models can generate commands for third-party applications, which expands their functionality. For example, they can create a command to search for airline tickets, which will be sent to the appropriate service. The ability to generate commands will become available in Yandex Cloud in the near future, which will open up new horizons for developers using models to interact with external applications.

Read materials on the topic:

«Авто.ру» made friends with a neural network: YandexGPT tells the story of a car

GosGPT: AI will be created in Russia that will translate the language of officials into ordinary language

Simplifies working with content: a text neural editor has appeared in Yandex Browser