Yandex B2B Tech, the corporate division of Yandex, has launched open-source visual-generative models (VLM) on its proprietary ML platform that simultaneously analyze images and text. With VLM models such as Deepseek VL2 Tiny and Gemma3 27B, companies will be able to create product descriptions from photos and quickly find the information they need in documents.

These models are available in Yandex Cloud AI Studio for batch processing of large numbers of images. The models can be used in a new mode: a huge number of requests can be sent to neural networks at once. For example, analyze user comments on social networks or compile a brief summary of many scientific articles. In total, about 20 large language (LLM) and VLM models are available for processing huge amounts of data.

According to the press service, Yandex now has one of the largest parks of open source neural networks in Russia - there are about 20 of them in total.

Pricing for LLM and VLM models when used on large amounts of data starts from 200,000 tokens (approximately 200 images or 360 pages of text). This use of models will cost half as much as in standard mode, and the result can be obtained within a day.

Among the already available models are Qwen2.5 and LLaMa 3.3, the reasoning neural networks QwQ and DeepSeek R1. As new models appear, they will almost immediately be deployed on the Yandex Cloud AI Studio platform. Soon, Yandex's VLM model, which is already used in Alice, Neuroexpert, Search and other company services, will also be available to customers. If a client needs to use the model for one-time requests, they can deploy the desired neural network on the cloud platform on dedicated resources.