PolarSeeker/OpenSeeker

What it solves

OpenSeeker addresses the lack of transparency and accessibility in frontier search agents. While many high-performing search agents exist, their training data is often proprietary. OpenSeeker provides a fully open-sourced set of training data and models to democratize the ability to build and deploy advanced search agents for complex information-seeking tasks.

How it works

OpenSeeker utilizes a ReAct-based search agent architecture. It provides fine-tuned models (v1 and v2) based on Qwen3-30B, which are trained using Supervised Fine-Tuning (SFT) on specialized datasets. The system integrates with external tools to perform search and information retrieval:

• Search and Visit Tools: Allows the agent to query the web and visit specific pages.
• E2B Sandbox: OpenSeeker-v2 incorporates E2B sandbox tools for enhanced capabilities.
• Evaluation Framework: Includes scripts to generate answers and evaluate performance across various benchmarks like BrowseComp and xbench.

Who it’s for

This project is designed for researchers and developers who want to build, evaluate, and deploy state-of-the-art open-source search agents without relying on closed-source training pipelines.

Highlights

• Fully Open-Sourced: Provides both the models and the full training data used to achieve high performance.
• SOTA Performance: OpenSeeker-v2 achieves state-of-the-art results among 30B-scale ReAct-based search agents on benchmarks such as BrowseComp and Humanity’s Last Exam.
• Tool-Augmented: Supports web search, page visiting, and sandbox execution via E2B.
• Academic Origin: The first work by a purely academic team to achieve SOTA search agent performance while open-sourcing all training data.