T-Technologies Open-Source Russian Model for Intelligent Document Search
T-Technologies has open-sourced T-Search, the first Russian-language model for multi-step search across companies' internal documents and databases. The solution speeds up and reduces the cost of deploying AI agents in business.
Unlike conventional search systems, T-Search gathers context using a specialized sub-agent that explores knowledge bases and builds a more complete context for the language model. The model is based on Qwen3.6-35B-A3B: of its 35 billion parameters, only 3 billion are active, allowing the solution to run on a single NVIDIA H100 GPU instead of sixteen. According to the developers, T-Search operates up to three times faster and costs up to five times less per token than open models of comparable quality, while deployment can reduce query processing costs by 20–100%.
The model has been published on Hugging Face under the Apache 2.0 license.
