Yandex B2B Tech Unveils Qwen3–235B‑A22B‑Instruct-2507, the Largest Language Model in the Russian Market

Yandex B2B Tech has launched access to the Qwen3–235B‑A22B‑Instruct-2507. According to Yandex, this is the largest language model available in the Russian market. The Qwen3–235B‑A22B‑Instruct-2507 operates in the cloud and, in some instances, outperforms top models from DeepSeek AI and OpenAI, as reported to Habr’s news service by the IT company’s press office. This model retains a significant amount of context, adeptly handling logical tasks and programming code.

The neural network features a contextual window of up to 256,000 tokens, allowing it to «memorize» vast amounts of information, which leads to more accurate and personalized responses. It supports 119 languages and dialects and has a rich knowledge base. In this version, the reasoning mode has been disabled, but the response quality is improved compared to the previous version, and the model operates more rapidly.

This model can be utilized for developing AI agents across various business sectors. For instance, it can aid in automating customer support by addressing common technical queries or function as a virtual assistant for online stores, assisting with product recommendations and handling returns. Alibaba has already begun integrating Qwen-based agents into its e-commerce services.

Deploying such models requires substantial computational resources and a dedicated engineering team. Businesses can access this neural network via API on the Yandex Cloud AI Studio platform.

The platform offers 24 models, including families like Qwen, DeepSeek, and Gemma. According to Arthur Samigulin, head of the product ML direction at Yandex Cloud, the company is customizing open-source models for local applications so that businesses can safely incorporate these neural networks into their systems and develop their own AI agents.

On June 24, 2025, VK released the RuModernBERT neural model for processing conversational Russian language into open access. RuModernBERT comprehends long texts in their entirety, without breaking them into segments. The model operates locally and does not rely on external APIs. According to VK, this helps reduce the burden on infrastructure.