Yandex has released a new generation of generative neural networks, YandexGPT 4. The line includes a powerful Pro language model and a lightweight Lite version. They surpass previous versions in response quality, and are capable of reasoning and processing queries four times longer.

The Pro version is more suitable for complex, time-consuming tasks, such as sales analysis. It can be used to sort emails and customer requests or analyze resumes. Lite will be useful in simpler scenarios where response speed is important.

Thus, in terms of the quality of answers, YandexGPT 4 Pro wins over the previous generation model in 70% of cases on average, and in some types of tasks — for example, in answers to open questions — it comes close to OpenAI’s GPT-4o . The Lite version responds no worse than Yandex’s most powerful previous generation model. All this became possible thanks to the improvement of the learning process.

The developers also taught them to reason. Before providing an answer, the AI ​​breaks down the task into simpler ones and solves them step by step, building a chain. This not only improves the quality of answers, but also helps analyze requests. For example, the new Pro model is able to study a customer complaint, understand the problem, and offer a solution.

At the same time, YandexGPT 4 models process four times more text, retain context longer, answer complex questions, analyze more information, make fewer mistakes and respond faster than third-generation models. They can also generate commands for external applications, such as searching for airline tickets. This feature will soon be available to third-party developers.

Fourth-generation models are already available via API in Yandex Cloud. Some options are currently in test mode. In the future, they will appear in Yandex services for a wider audience. The first such service will be Alice with the Pro option enabled.