Alibaba Launches Open-Source AI Agent, Citing Efficiency and Performance Against OpenAI
- tech360.tv
- Sep 19
- 2 min read
Alibaba Group Holding has unveiled a new open-source artificial intelligence agent, which it states matches the performance of OpenAI’s flagship Deep Research tool while offering greater efficiency. The company describes its deep research agent as a “leading open-source deep research” tool.
Alibaba stated its agent showed "incredible efficiency" compared with US proprietary tools because it has only 30 billion parameters. This is significantly fewer than the estimated parameter counts of the models that power US deep research agents.
The new agent achieved industry-leading scores across various advanced benchmarks, including Humanity’s Last Exam, according to a graphic released by Alibaba. This challenging set of academic questions tests the limits of existing AI systems.
Alibaba's agent scored 32.9% on this benchmark, surpassing OpenAI’s Deep Research, which achieved 26.6% earlier this year. Adina Yakefu, a machine learning community manager at open-source platform Hugging Face, described Alibaba’s self-reported benchmark scores as “amazing.”
The agent has been integrated into Alibaba’s maps application, Amap, and its AI-powered legal research tool, Tongyi FaRui. Users of Amap can leverage the agent’s web retrieval capabilities to plan multi-day trips.
Tongyi FaRui has been updated with the agent’s research functions, enhancing its ability to retrieve case law with verified citations, according to a blog post on Tuesday by Alibaba’s AI search development team, Tongyi Lab.
Deep research agents are AI tools designed to perform complex, multi-step web retrieval tasks. OpenAI launched the first such agent, Deep Research, integrating it into ChatGPT in Feb. Other major US technology organisations, including Google DeepMind, have also introduced similar tools.
Parameters are variables that encode an AI model’s “intelligence” and are adjusted during the training process. Generally, a higher number of parameters indicates a more powerful model, but it also requires more computational resources to train and operate.
The strength and efficiency of Alibaba’s agent stem from its innovative data curation pipeline, which produced “very high-quality” synthetic training data, according to Tan Sijun, an AI researcher at the Sky Computing Lab of University of California, Berkeley.
Synthetic training data is generated by AI systems instead of being sourced from the real world. As real-world data becomes increasingly scarce, AI companies are turning to synthetic data to train new systems.
Alibaba’s synthetic data solution was applied throughout the entire training pipeline and incorporated a new technique enabling a “data flywheel.” This process reuses data generated during training to enhance the model without human intervention.
Its developers wrote that this approach “ensures both exceptional data quality and massive scalability, breaking through the upper limits of model capabilities.” They noted this training pipeline has yet to be validated on base models significantly larger than 30 billion parameters.
The company also stated that the agent’s context length of 128,000 tokens remains a limitation for many complex research tasks that require long inputs.
Alibaba Group Holding released an open-source AI agent that rivals OpenAI’s Deep Research tool in performance.
The new agent claims "incredible efficiency," operating with 30 billion parameters, fewer than estimated counts for US competitors.
It achieved a 32.9% score on Humanity’s Last Exam, surpassing OpenAI’s 26.6% score obtained earlier in the year.
Source: SCMP
