Introduction
TIGER-AI-Lab has released OpenResearcher, a fully open pipeline for training deep research agents — AI systems that answer complex questions by searching, browsing, and reasoning across many steps rather than answering from memory. The release, described in the paper "A Fully Open Pipeline for Long-Horizon Deep Research Trajectory Synthesis", goes well beyond a single model checkpoint: it publishes the training recipe, a dataset of roughly 96,000 research trajectories, training code, an evaluation framework, and an offline search environment for reproducing the whole pipeline end to end.
The headline number is the scale of the training data itself: TIGER-AI-Lab labels the release 96K long-horizon research trajectories (the paper's precise count is just over 97,000 sequences), many containing more than 100 tool calls apiece, built specifically to teach a model how to research rather than what to answer.
What a research trajectory teaches the model
Most instruction-tuned models answer a question directly from parameters learned during pretraining. OpenResearcher instead trains agents to work through a research loop:
- Search for relevant documents
- Open promising pages
- Find the specific fragments that answer the question
- Gather evidence across multiple sources
- Reason over several steps before committing to an answer
- Produce a final answer grounded in what was retrieved
Each of these trajectories is a full transcript of an AI agent working through this loop on a real research question, generated by using GPT-OSS-120B as a teacher model to explore an offline collection of 15 million documents — roughly 11 billion tokens — through exactly those three tools — search, open, and find.
Why an offline search environment
Training research agents against live web search APIs is slow, costly, and non-reproducible — results change over time and rate limits throttle large-scale trajectory generation. OpenResearcher sidesteps this by separating dataset construction from retrieval: the entire pipeline runs against a static, offline document index, so generating trajectories, fine-tuning on them, and evaluating the result are all cheap and deterministic. That same offline environment lets other researchers rerun the exact experiments and test their own filtering strategies or retrieval methods without needing API budget.
Results
The team used the trajectory dataset to fine-tune OpenResearcher-30B-A3B, a mixture-of-experts model with 30 billion total parameters and about 3 billion active per forward pass (the "A3B" in the name), via supervised learning. On BrowseComp-Plus, a benchmark for deep research agents, the fine-tuned model reached 54.8% accuracy — a 34.0-point improvement over its base model — while holding up on related benchmarks, including BrowseComp, GAIA, and xbench-DeepResearch.
What's released
Everything is public:
- OpenResearcher-Dataset — the ~96K research trajectories, on Hugging Face
- OpenResearcher-Corpus — the offline 15-million-document, ~11-billion-token collection used for retrieval
- Training code and the distillation recipe used to generate trajectories from GPT-OSS-120B
- Evaluation framework and logs for reproducing the benchmark results
- Trained model weights for the 30B-A3B checkpoint
The repository is available at github.com/TIGER-AI-Lab/OpenResearcher.
Conclusion
OpenResearcher's contribution isn't a single leaderboard number — it's making the entire recipe for training a deep research agent reproducible, from raw document corpus through trajectory synthesis to the fine-tuned model. That matters for a category where most strong systems (commercial deep-research products from major labs) keep their training data and search infrastructure closed. By publishing the data, the offline environment, and the code together, TIGER-AI-Lab gives other teams a concrete baseline to fine-tune, extend, or audit rather than a black box to reverse-engineer.