T-Bank Releases 32 Billion Parameter LLM to the Public

T-Bank has made two large language models (LLM), T-Pro and an updated version of T-Lite, publicly accessible. This was reported by Vedomosti, citing a bank representative.

The professional version of the neural network boasts 32 billion parameters, while T-Lite has 7 billion. The latter was introduced in July 2024.

The number of parameters indicates a model’s ability to consider more context, better retain information, and provide accurate answers. For comparison, Meta’s Llama 3.1 405B has 405 billion parameters, and the French startup Mistral’s Large 2 has 123 billion.

Open access to T-Pro and T-Lite will allow businesses to use neural networks for internal tasks and create new products based on the developed models at no cost. Vedomosti cited examples such as smart chatbots for support services and assistants for employees, like tools for coding, report writing, or research.

The models are part of the LLM Gen-T family, designed for specialized tasks, distinguishing them from the more versatile ChatGPT.

T-Lite and T-Pro are built on Qwen-2.5 from Chinese developer Alibaba Group and are adapted for the Russian language.

“This approach allows T-Technologies to significantly reduce the costs of creating large language models—by 80–90% compared to companies that train them from scratch. Meanwhile, the quality of LLM products in Russian surpasses solutions based on open and proprietary models from OpenAI, Google, and Anthropic,” noted the bank representative.

Businesses can deploy language models on their servers and further train them for specific cases, emphasized Dbrain CEO Alexey Khakhunov.

In December, a virtual AI-based assistant for psychologists was introduced in Russia, doubling productivity, service quality, and revenue.

In the same month, Russian scientists presented a new open environment, XLand-MiniGrid, designed for developing reinforcement learning algorithms—a distinct area in AI research.