TLDR: Alibaba’s AI division, Tongyi Lab, has officially open-sourced WebSailor, a new web AI agent designed for highly complex information-seeking tasks. Released on July 4th, 2025, WebSailor has achieved state-of-the-art performance on challenging benchmarks like BrowseComp and GAIA, outperforming many existing open-source and even some proprietary models. This strategic move underscores Alibaba’s commitment to fostering collaboration and accelerating innovation within the global AI ecosystem.
Alibaba’s Tongyi Lab has made a significant stride in the artificial intelligence landscape with the official open-source release of WebSailor, a sophisticated web AI agent. Announced on July 7, 2025, following its initial release on July 4, WebSailor is engineered to tackle extremely complex information retrieval and reasoning challenges, positioning itself as a frontrunner in the field.
WebSailor has demonstrated remarkable capabilities, achieving top scores on critical benchmarks. On the BrowseComp evaluation set, it scored 12.0% in the English version and an impressive 30.1% in the Chinese version. Furthermore, it attained a 55.4% score on the GAIA benchmark. These results highlight WebSailor’s advanced ability to navigate, comprehend, and interact with intricate web environments. Notably, the WebSailor-32B and WebSailor-72B versions have surpassed several prominent closed-source models, including DeepSeek R1 and Grok-3, and are reported to trail only behind OpenAI DeepResearch in certain evaluations. This performance underscores its capacity for ‘superhuman’ reasoning, particularly in scenarios characterized by high uncertainty.
At the core of WebSailor’s exceptional performance is its innovative training methodology. The agent is trained to address ‘Level 3’ problems—tasks that lack clear solution paths and necessitate deep, exploratory reasoning. To facilitate this, Tongyi Lab developed SailorFog-QA, a unique data synthesis pipeline. This system constructs complex knowledge graphs from real-world websites and deliberately obfuscates information, compelling the agent to reason and connect disparate facts. The training process meticulously avoids ‘stylistic contamination’ by reconstructing concise, action-oriented thoughts from expert trajectories. It employs a two-stage training regimen: an initial ‘cold start’ using Rejection Sampling Fine-Tuning (RFT) to build foundational skills, followed by reinforcement learning powered by Duplicating Sampling Policy Optimization (DUPO), an efficient agentic RL algorithm.
Alibaba’s decision to open-source WebSailor aligns with a growing strategic trend in the AI industry. Companies are increasingly recognizing the benefits of open-sourcing, viewing it as a catalyst for innovation rather than a competitive disadvantage. This approach fosters transparency, builds trust within the developer community, encourages collaborative model improvement, and expands potential applications beyond what a single organization could achieve. For major Chinese technology firms like Alibaba, open-sourcing also serves as a strategic move to counterbalance Western AI dominance and accelerate the global adoption of their technological frameworks and standards.
Also Read:
- Baidu Pivots to Open-Source AI with Ernie 4.5 Release, Intensifying Global Competition
- Alibaba Cloud Commits $60 Million to Accelerate Global AI Adoption via Enhanced Partner Ecosystem
The construction plan and select datasets for WebSailor are now accessible on GitHub, inviting developers and researchers worldwide to explore and build upon this groundbreaking technology. This move is anticipated to significantly accelerate advancements in web agent technology, opening new avenues for automating online tasks and enhancing efficient information gathering across various sectors, including e-commerce, customer service, and market research. WebSailor represents a progression from Alibaba’s earlier projects, WebDancer and WebWalker, further solidifying the company’s contributions to the evolving landscape of web agents.
