T-Tech Open-Sources T-Search: High-Performance Agentic Retriever for Multi-Step Search That Runs on a Single GPU

The open-source AI community just received a major efficiency upgrade. T-Tech has released T-Search, an agentic retriever purpose-built for difficult multi-step search tasks in both English and Russian. 

The model is based on Qwen3.6-35B-A3B (a Mixture-of-Experts architecture) and delivers state-of-the-art retrieval quality while dramatically lowering the hardware requirements that previously defined this level of performance.

---

Why T-Search Matters: Economics First

The standout feature of T-Search is its extraordinary cost-performance ratio.

• The underlying MoE model activates only ~3 billion parameters per token despite having 35 billion total parameters.
• This allows high-quality inference on a single GPU — something that previously required clusters of 16 GPUs for comparable retrieval performance.
• As a result, the cost of running sophisticated search agentics drops by roughly an order of magnitude, while quality continues to improve.

---

How the Agent Works

T-Search is not a simple embedding + reranker pipeline. It functions as a true agent that plans, executes, and reflects across multiple rounds using a ReAct-style loop.

Key technical innovations include:

• Multi-round operation with a configurable search budget.
• Compact “memory” transfer between rounds — only essential evidence chunks, coverage state, and next steps are carried forward. Full tool history and noise are discarded.
• This design keeps the context window comfortably within ~32K tokens per round (up to 65K supported in serving).
• Tools available to the agent: `search_corpus`, `save_and_advance`, and `finalize_ranking`.

The result is an efficient, controllable search process that can be tuned from “fast and cheap” (single rollout) to “maximum quality” (multiple parallel rollouts).

Performance That Beats Much Larger Models

Evaluated on the official T-Search harness across English and Russian benchmarks (including TRuST, SynthComp, and adapted versions of BrowseComp and SealQA), T-Search sets new standards for open models:

Key takeaways:

• A single rollout already surpasses several much larger models.
• Scaling to three parallel rollouts + Reciprocal Rank Fusion (RRF) boosts performance from 55.96 → 61.33.
• The model shows strong robustness across different retriever backends (embedding models + optional LLM reranking).

Notably, T-Tech reports no meaningful gains on general-purpose benchmarks — the specialization is laser-focused on retrieval agentics.

---

Fully Open Source Under Apache 2.0

Everything needed to reproduce and build upon the work is released:

• Main model: `t-tech/T-Search`
• Quantized versions: `t-tech/T-Search-FP8` and `t-tech/T-Search-NVFP4`
• Harness: https://github.com/turbo-llm/t-search-harness (the exact inference framework used for training and evaluation)
• Datasets: TRuST, SynthComp (English & Russian), and others — available on Hugging Face

Collection: https://huggingface.co/collections/t-tech/t-search

The model was trained entirely on synthetic search trajectories generated with the same harness used at inference time. This closed-loop approach enabled high-quality supervised fine-tuning followed by reinforcement learning (using recall-based rewards).

---

Bottom Line

T-Search proves that specialized agentic training + smart architecture choices can deliver better results than simply scaling model size. By combining a lightweight MoE backbone, efficient multi-round memory management, and flexible rollout scaling, T-Tech has made high-quality multi-step search dramatically more accessible.

For researchers, developers, and companies building RAG systems, search agents, or knowledge-intensive applications — especially those working with English and Russian — T-Search represents one of the most compelling open releases of 2026.

Try it today:  
→ Collection: https://huggingface.co/collections/t-tech/t-search
→ Harness: https://github.com/turbo-llm/t-search-harness

The era of expensive, cluster-only retrieval agents may be coming to an end.

---

Also read:

• A Defining Moment for British Media: Sky’s £1.6bn Acquisition of ITV’s Broadcast Business
• ACEA Responds to Proposed CBAM Extension Discussions for Autos
• OPEC Lowers 2026 Oil Demand Growth Forecast to 780,000 bpd
• Character AI Enters the Microdrama Market: From Chatbots to AI-Generated Series
• How Aluminum Unibody, Vapor Chamber, and 48MP Cameras Shape iPhone 17 Pro

 

---

Thank you!