Т-Банк выпустил T-Search: open-source агент для многошагового поиска на русском

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

How T-Search differs from ordinary search

T-Search does not write a final answer — instead, it returns a ranked list of relevant chunks, i.e., text fragments. Generating the final answer on top of this list can be entrusted to any model, depending on the requirements of a specific product. This is a key architectural decision: the retriever and the generator are separated, and the developer chooses which LLM formulates the answer for the user.

The model is tailored for multi-step search — when the answer cannot be found with a single query and instead requires sequential refinement, combining sources, and filtering out irrelevant information.

• 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 ready-made text
• Part of the open Gen-T lineup (after T-Lite and T-Pro)
• Released alongside the model in open source — 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. For 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 no junk makes it into the training set.

The approach to synthetic data here is a matter of principle: 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 doesn't write the final answer — it returns a ranked list of chunks, and generation can be entrusted to any model depending on the requirements of a specific product," —

Tolya Potapov, MLE at T-Bank, in the company's blog on Habr.

Why the TRuST and SynthComp benchmarks are needed

Alongside the model, T-Bank released two benchmarks for measuring the quality of agentic search. TRuST, according to the team's description, is the first benchmark for complex agentic search in Russian. SynthComp is a synthetic benchmark in Russian and English with fixed indexes, which allows reproducible comparison of different retrievers on the same data.

Open benchmarks in Russian matter in their own right: most metrics for retrieval and RAG have historically been tailored to English, and until now there was almost no way 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, a specialized retriever has now appeared, along with tools for evaluating it. For teams building RAG and search agents in Russian, this is a ready-made open-source component and a way to honestly measure its quality.

Frequently asked questions

What is T-Search?

T-Search is T-Bank's open-source agent retriever for complex multi-step search, based on Qwen3.6-35B-A3B. It does not generate the final answer but returns a ranked list of relevant text fragments, on top of which any chosen model forms 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 took place in two stages — SFT and RL.

What benchmarks did T-Bank release?

Alongside the model, two benchmarks were released in open source: TRuST — the first benchmark for complex agentic search in Russian, and SynthComp — a synthetic benchmark in Russian and English with fixed indexes.