t-tech/T-Search
T-Search is a 35.1 billion parameter agentic retriever model developed by t-tech, built upon the Qwen3.6-35B-A3B architecture. It is specifically designed for difficult multi-step search tasks in both English and Russian, planning and executing searches across multiple rounds. This model excels at agentic retrieval, outperforming its base model and larger open models on various retrieval benchmarks, and maintains strong retrieval quality across different retriever configurations.
Loading preview...
T-Search: An Agentic Retriever for Multi-Step Search
T-Search is a 35.1 billion parameter agentic retriever model developed by t-tech, built on the Qwen3.6-35B-A3B architecture. It is engineered for complex multi-step search tasks in English and Russian, operating as an agent that plans and executes searches across multiple rounds with a configurable budget. The model's primary function is to collect evidence by formulating queries, inspecting retrieved snippets, and returning a ranked set of evidence chunks for downstream processing.
Key Capabilities
- Agentic Retrieval: T-Search issues queries, inspects results, tracks coverage, and carries compact evidence and search state across rounds, ultimately returning a ranked chunk set.
- Enhanced Retrieval Quality: It consistently outperforms its base model and other larger open models on average across evaluated retrieval benchmarks.
- Retriever Robustness: The model demonstrates strong retrieval quality across various retriever configurations, having been trained and evaluated with multiple backend retrievers.
- Multi-stage Training: T-Search was trained in two stages: supervised fine-tuning on 11K synthetic search trajectories per language, followed by reinforcement learning with GSPO and recall-based rewards.
- Benchmark Performance: Achieves an average Recall@10 of 55.96 (N=1) and 61.33 (N=3) across benchmarks like TRuST, SynthComp-En/Ru, BrowseComp-Plus, and SealQA, significantly outperforming its Qwen3.6-35B-A3B baseline (41.54).
Good for
- Complex Information Retrieval: Ideal for applications requiring multi-step search and evidence collection from a corpus.
- English and Russian Search: Optimized for robust performance in both English and Russian language contexts.
- Integrating with Search Backends: Designed to work with a corpus search backend, formulating queries and processing retrieved snippets.
- Building Advanced RAG Systems: Provides a strong foundation for systems that require sophisticated evidence gathering before generation or reranking.