Built on Qwen3.6-35B-A3B and trained on synthetic data, it returns a ranked set of evidence chunks rather than a final answer, leaving answer generation to a separately chosen model. The release also includes TRuST, a hand-built Russian benchmark for complex agentic search, and SynthComp, a synthetic benchmark in Russian and English. Its harness uses bounded search rounds and explicitly preserved evidence to manage context growth.
Raw SignalSep 28, 20261 min read1 source
Llm deployment strategy and automation economics
In one minute
T-Bank has released T-Search, an open agentic retriever for complex multi-step search.