T-Tech reports higher average Recall@10 than far larger open models — and its best score comes with three parallel rollouts.
T-Tech has released T-Search into the open, putting the model, its harness and the datasets under Apache 2.0. The system is a search agent that resolves a question over several retrieval passes, in Russian and English.
It is built on Qwen3.6-35B-A3B, and a related collection has been published on Hugging Face.
The central claim is hardware. T-Search is a mixture-of-experts system that runs on a single GPU; reaching comparable quality previously meant standing up clusters of 16 GPUs.
T-Tech puts T-Search ahead of much larger open models on average Recall@10. The models named are Qwen3.5-397B, GLM-5.2 and Kimi-K2.6, and the comparison is drawn on that metric alone.
Context handling is where the agent's design shows. Instead of carrying the full history through every step, it works in rounds, passing only a condensed memory forward — so the context stays small even as a search stretches over several hops.
How hard it searches is left to the operator. At one end of the dial is a single fast, cheap search. At the other, three rollouts run in parallel and their results are merged with RRF (Reciprocal Rank Fusion).
That dial is also where the strongest number sits. Three parallel rollouts fused with RRF take average Recall@10 from 55.96 to 61.33; the single fast search is the cheaper end of the same range.