Frontier AI agents are usually the exclusive domain of tech giants with bottomless budgets. But a new open-source model proves that you don't need billions in compute to build world-class search capabilities. You just need the right data.

Researchers have developed OpenSeeker-v2, a search agent that challenges the industry standard. Typically, building such powerful tools requires a massive, resource-heavy pipeline involving pre-training, continual pre-training, supervised fine-tuning, and reinforcement learning. This new approach skips the complexity. By focusing on high-difficulty training trajectories and strict data filtering, the team achieved state-of-the-art results using only supervised fine-tuning on a modest dataset of 10,600 examples.

The results are striking. OpenSeeker-v2 outperforms Tongyi DeepResearch, a model trained with far more resources, across four major benchmarks. It scored 46.0% on BrowseComp, 58.1% on BrowseComp-ZH, 34.6% on Humanity's Last Exam, and 78.0% on xbench. The key was not more compute, but smarter data synthesis. The team scaled knowledge graphs for richer exploration, expanded tool sets for broader functionality, and applied strict low-step filtering to ensure efficiency.

This breakthrough democratizes access to frontier search technology. For the first time, an academic team has created a top-tier search agent within its scale and paradigm using only supervised fine-tuning. By open-sourcing the model weights and findings, the researchers are proving that simple, well-curated data can rival industrial might, making advanced AI research accessible to everyone.