T-Pro 2.0

T-Pro 2.0 is a large language model from the T-Technologies team designed for analytics-heavy tasks and process automation. Built on Qwen-2.5, it targets Russian-language queries with optimized tokenization and a hybrid reasoning mode to reduce compute costs.

What it’s used for

T-Pro 2.0 is positioned for scenarios where you need both fast responses and multi-step reasoning, including building AI agents for business workflows.

• Breaks complex tasks into logical steps
• Generates and checks hypotheses
• Produces answers in Russian
• Supports fine-tuning on company data
• Can run tasks with minimal human involvement

The model is reported to handle up to 45% of customer requests at T-Bank without human operators. It also claims lower cost for Russian queries compared with Qwen3 and DeepSeek R1-Distil, and strong results on Russian benchmarks such as MERA and ruMMLU.

Access and deployment

T-Pro 2.0 is available for free on Hugging Face under the Apache 2.0 license for download and integration. It runs in Python (Transformers) and typically requires an NVIDIA GPU environment. Cloud.ru pricing is listed as 85 ₽ per million input tokens and 170 ₽ per million output tokens. The model supports Russian and other Cyrillic languages.

Key specs

• 32B parameters
• T-Wix dataset (500K examples)
• Compatible with SGLang and OpenTX