On November 29, StepFun, a unicorn in the realm of large models, made its GUI Agent technology, the GELab-Zero Suite—similar in nature to Doubao Mobile Assistant—available as open-source. Alongside this, the company also open-sourced a 4B GUI Agent model (GELab-Zero-4B-preview), complete with a full suite of supporting infrastructure. This model has comprehensively set new benchmarks for performance among models of its size, excelling across multiple GUI evaluation metrics on both mobile and computer platforms, and achieving SOTA (State-of-the-Art) results. Additionally, StepFun has open-sourced its proprietary evaluation standard, AndroidDaily, which is grounded in real-world business scenarios. This move is aimed at steering the development of model evaluation in the GUI field towards consumer-grade and large-scale applications.

Now, enterprise users and developers have the opportunity to experience GELab-Zero firsthand via the GitHub and HuggingFace platforms. The model is capable of operating at the 4B scale on consumer-grade hardware, striking a balance between low latency and privacy. It offers one-click deployment across multiple terminals, facilitating distribution to numerous mobile phones. Moreover, it records interaction trajectories for enhanced observability and reproducibility, and also supports multi-modal agent paradigms.