T-Bank releases T-Search, an open-source agent for multi-step search in Russian

In July 2026, T-Bank released T-Search, an open-source retriever agent for complex multi-step search. It is built on Qwen3.6-35B-A3B and trained entirely on synthetic data. It is a new class of model in the open Gen-T lineup, which already includes the Russian-language T-Lite and T-Pro families.

How T-Search differs from conventional search

T-Search does not write the final answer. It returns a ranked list of relevant chunks, meaning text fragments. Any model can generate the final answer from that list, depending on the requirements of a specific product. This is the key architectural decision: the retriever and the generator are separate, and the developer chooses which LLM words the answer for the user.

The model is built for multi-step search, where a single query can't find the answer and the system has to refine queries step by step, combine sources, and filter out irrelevant material.

• Base model: Qwen3.6-35B-A3B (MoE architecture)
• Training: entirely on synthetic data, built for T-Bank's own harness
• Model output: a ranked list of chunks, not finished text
• Part of the open Gen-T lineup (after T-Lite and T-Pro)
• Open-sourced alongside the model: two benchmarks, TRuST and SynthComp

How the model was trained

T-Search was trained in two stages, SFT (supervised fine-tuning) and RL (reinforcement learning), on data that T-Bank generated itself. To do this, the team built a "factory" of synthetic search tasks with an adversarial verification loop: tasks are created and immediately checked for quality, so that junk doesn't end up in the training set.

The synthetic approach is a deliberate choice here. Real labeled datasets for complex agentic search in Russian are practically nonexistent, so the training data had to be generated for the specific harness rather than collected by hand.

"T-Search does not write the final answer.

It returns a ranked list of chunks, and generation can be entrusted to any model that fits the requirements of a specific product," said Tolya Potapov, an MLE at T-Bank, in the company's blog on Habr.

Why the TRuST and SynthComp benchmarks matter

Alongside the model, T-Bank open-sourced two benchmarks for measuring the quality of agentic search. According to the team, TRuST is the first benchmark for complex agentic search in Russian. SynthComp is a synthetic benchmark in Russian and English with fixed indexes, which allows different retrievers to be compared reproducibly on the same data.

Open Russian-language benchmarks matter in their own right: most retrieval and RAG metrics have historically been built for English, and until now there was almost nothing against which to objectively measure the quality of Russian-language search.

What this means

T-Bank continues to build out the open Russian AI ecosystem. After the T-Lite and T-Pro language models, it now has a specialized retriever plus tools to evaluate it. For teams building RAG systems and search agents in Russian, this is a ready-made open-source component and a way to measure its quality honestly.

FAQ

What is T-Search?

T-Search is T-Bank's open-source retriever agent for complex multi-step search, built on Qwen3.6-35B-A3B. It does not generate the final answer. It returns a ranked list of relevant text fragments, from which any chosen model builds the answer.

What data was T-Search trained on?

The model was trained entirely on synthetic data that T-Bank generated for its own harness through a "factory" of search tasks with an adversarial verification loop. Training ran in two stages, SFT and RL.

Which benchmarks did T-Bank release?

Two benchmarks were open-sourced alongside the model: TRuST, the first benchmark for complex agentic search in Russian, and SynthComp, a synthetic benchmark in Russian and English with fixed indexes.