OpenSeeker Challenges OpenAI and Alibaba with Its AI Model

OpenSeeker: An Open-Source Alternative to AI Giants

OpenSeeker presents itself as a fully open-source AI search agent, designed to operate autonomously on the web at various levels. What sets OpenSeeker apart is its complete transparency: all training data, source code, and model weights are publicly available. The stated goal is ambitious: to end the data monopoly held by major companies such as OpenAI or Alibaba.

The OpenSeeker model is trained on questions that emerge from the structure of web links, a method that goes beyond simple searches. Through a teacher-student learning process, the model can extract relevant information even from noisy data.

Despite limited training resources, OpenSeeker manages to compete with much more resource-intensive systems, although it still cannot match the most advanced proprietary models.

With only 11,700 training data points and a single training session, OpenSeeker achieves results comparable to giants like Alibaba. All data, code, and the model are freely accessible, marking a significant step towards a more open AI.

The Data Monopoly in AI Search Agents

Powerful AI search agents, capable of navigating and collecting information on the internet autonomously, have long been the domain of large tech companies. Players like OpenAI, Google, and Alibaba guard their training data jealously. Even when some projects share their model weights, the underlying data often remains inaccessible.

This strict control of data has slowed the progress of open research for nearly a year, according to researchers from Shanghai Jiao Tong University. With OpenSeeker, these researchers aim to reverse this trend by making all training data, code, and model weights available under an MIT license.

An Innovative Approach Based on Web Link Structure

OpenSeeker relies on two key concepts to generate its data. For question-answer pairs, the team uses the actual structure of web links as a starting point, generating questions from this base. The system begins with randomly chosen starting pages from a web corpus of about 68 GB of English data and 9 GB of Chinese data, then follows hyperlinks to related pages to extract crucial information.

Specific names and terms are then replaced with vague descriptions, preventing a simple keyword search from providing the answer. This necessitates genuine research and multi-step reasoning.

A two-step filtering process eliminates unusable questions: a solid base model should not be able to answer without tools but must be capable of resolving questions with complete context. If either condition fails, the question is rejected.

The second idea focuses on the search paths that the model learns. Web pages often contain noise that can degrade the quality of the recorded solution paths. During data generation, a teacher model receives a cleaned summary of previous search results to make better decisions.

During training, the student model sees the raw, uncleaned data but must reproduce the high-quality decisions of the teacher. This forces it to learn to distinguish the signal from the noise on its own.

Data Quality Takes Precedence Over Quantity

OpenSeeker is based on the Qwen3-30B-A3B model and was trained with only 11,700 data points in a single session, using supervised fine-tuning without reinforcement learning or repeated adjustments.

According to the article, the model achieved a score of 48.4% on the BrowseComp-ZH benchmark in Chinese, surpassing Alibaba's Tongyi DeepResearch model, which scored 46.7%. Alibaba's model underwent a three-step process of extended training, supervised fine-tuning, and reinforcement learning.

On OpenAI's English BrowseComp benchmark, OpenSeeker scores 29.5%, nearly double the 15.3% of DeepDive, the previous leader among fully open agents.

A comparison with MiroThinker highlights the importance of data quality over raw quantity: this model was fed 147,000 training examples but only reaches 13.8% on BrowseComp-ZH. OpenSeeker achieves a score 3.5 times higher with less than one-twelfth of the data.

OpenSeeker's training data in Chinese requires an average of 46 tool calls per task, compared to only 27 for BrowseComp-ZH.

However, there remains a gap with the most powerful proprietary systems. OpenAI's GPT-5-High scores 54.9% on BrowseComp, and DeepSeek-V3.2 with 671 billion parameters reaches 51.4%. OpenSeeker operates with a fraction of the model size and training effort.

The issue of access to high-quality training data has been a central concern in the AI industry for some time. Last year, a research team released the Common Pile, a text dataset of 8 TB built from open-licensed sources. So far, this has not significantly contributed to shaking the dominance of commercial models.