Yandex Cloud AI Studio

The main articles are:

• What is IaaS
• Big Data
• LLM (Large Language Models)

2025

Access to neural networks gpt-oss-120b and gpt-oss-20b

Yandex B2B Tech has opened up access to OpenAI's new reasoning neural networks gpt-oss-120b and gpt-oss-20b. Open source models can be used via API on the Yandex Cloud AI Studio platform. These models will help the business create agency systems - for example, to automate recruitment and technical support, analyze and process documents and primary communication with customers. Yandex B2B Tech announced this on August 7, 2025.

The availability of these API models solves the problems of Russian business in integrating OpenAI technologies into their own business processes. Deploying open source neural networks on your own infrastructure can be difficult, since this requires a significant amount of computing power. To use models directly from the developer, you need to transfer data for processing abroad, which may not meet the requirements of Russian legislation. When using Yandex Cloud AI Studio, data is stored and processed in Russian data centers, while the service meets all the requirements of the Law "On Personal Data."

These open source neural networks are comparable in quality to the leading models OpenAI o3-mini and o4-mini, and in some scenarios they surpass GPT-4o and o1. In new models, you can adjust the intensity of reasoning and the rate of response generation. Calling features to interact with external applications will soon be available for these models on the cloud platform. This will help, for example, find the information you need on the Internet when generating a response.

Access to Qwen3-235B-A22B-Instruct-2507

Yandex B2B Tech has opened access to a Qwen3-235B-A22B-Instruct-2507 model that holds a large amount of context, solves logical problems with high quality and works with code. Business will be able to use the neural network to develop AI agents to automate business processes. Yandex B2B Tech announced this on July 25, 2025.

Qwen3-235B-A22B-Instruct-2507 is a neural network with a large context window of up to 256 thousand tokens. The model can "hold" a large amount of information in memory, so it is more personalized and answers questions accurately. In addition, it supports 119 languages ​ ​ and dialects and has a large knowledge base. In this version of the model, the reasoning mode is turned off - when a neural network stepwise solves complex problems. At the same time, in terms of the quality of the answers issued, it is ahead of the previous version and generally works faster.

The updated model is applicable to the creation of AI agents in different areas of business. For example, you can create an agent to automate support so that you can quickly solve common technical problems for users. The virtual assistant for the online store will help customers find products and automate the return system. Alibaba is already integrating Qwen-based agents into its e-commerce services to serve customers.

For many companies, it is difficult to deploy models of this size on their own. This requires large computing resources and an engineering team that can make the inference fast. On the Yandex Cloud AI Studio platform, a business can use a neural network via API. The cost of the model will be 50 kopecks per 1000 tokens.

We flexibly adapt open source models to local features and tasks so that companies can safely integrate neural networks into their web applications and develop their own AI agents, "said Artur Samigullin, head of the Yandex Cloud product ML department.

24 models are available on the Yandex Cloud AI Studio platform for July 2025.

Opening access to open visual-generative models

Yandex B2B Tech has opened access to open source visual-generative models (VLMs) such as Deepseek VL2 Tiny and Gemma3 27B. The company announced this on April 24, 2025. New technology solutions allow companies to analyze images, compile product descriptions and process large amounts of visual information.

About 20 large language and visual models are available in Yandex Cloud AI Studio. Among them are Deepseek VL2 Tiny, Qwen2.5 VL, Gemma3 27B and other advanced opensource solutions for image analysis.

𝓲

Фото: Yandex

Yandex Cloud CTO Dmitry Andreev noted the advantages of the new platform. Business gets the opportunity to test various neural networks to solve specific problems, further train models for specific needs and run them with minimal programming.

Key capabilities of the new neural networks include classifying goods by category, compiling descriptions for each image, finding and identifying defects in photographs, analyzing the design and functionality of interiors.

According to the company, charging models starts from 200 thousand tokens (approximately 200 images or 360 pages of text). The cost of using in batch mode will be half the standard tariff, and the result can be obtained during the day.

Among the available models are Qwen2.5, Llama 3.3, reasoning neural networks QwQ and DeepSeek R1. The company plans to quickly deploy new opensource models on the platform as they become available.

In the near future, customers will also have access to Yandex's own VLM model, which is already used in services such as Alice, Neuroexpert and Search. Customers will be able to deploy the necessary neural network on a cloud platform with dedicated resources for one-time requests.^[1]

Notes