Skywork-DeepResearch V2 represents a breakthrough in deep research agents, combining a novel data synthesis piepline and an end-to-end reinforcement learning pipeline with cutting-edge techniques for verification and reward shaping. Our approach features:

• Advanced RL Pipeline: Non-symmetric verification coupled with dense, clue-aware reward shaping
• High-Quality Synthetic Data: Built through a two-stage process that iteratively enhances the quality, diversity, and complexity of the data
• Efficient Infrastructure: Parallel inference capabilities and self-learning multi-agent loops enable rapid continual improvement
• State-of-the-Art Performance: Sets new benchmarks on demanding web-research tasks, outperforming Claude-4-Opus, GLM-4.5, and other leading systems

Skywork-DeepResearch V2 is now available through API access.

Skywork-DeepResearch V2 demonstrates exceptional performance on the BrowseComp benchmark, establishing new state-of-the-art results. The following evaluation shows our model significantly outperforming strong baselines including Claude-4-Opus, GLM-4.5, and other leading systems.

With parallel thinking enabled, Skywork-DeepResearch V2 achieves 38.7% accuracy, surpassing the previous state-of-the-art (Grok-4) by an impressive 6.1 percentage points.

Skywork-DeepResearch V2 is currently available through API access. To request access, please submit an application to [email protected] with the following information:

• Affiliated Institution: Your organization or academic institution
• Intended Usage: Detailed description of your use case and research objectives
• Expected Maximum Concurrent Requests: Estimated API load requirements
• Contact Information: Primary contact details for technical correspondence
• Additional Requirements: Any specific needs or constraints for your implementation

Our team will review your application and provide access credentials along with API usage instructions upon approval.

For questions, support, or general inquiries about Skywork-DeepResearch V2, please contact us at [email protected].

If you find our work helpful or use it in your research, please consider citing:

@misc{skywork2025deepresearch,
  title = {Skywork-DeepResearch},
  author = {Skywork AI},
  howpublished = {\url{https://github.com/SkyworkAI/Skywork-DeepResearch}},
  year = {2025},
}