Step-GUIEdge is a pioneering open-source, edge-side GUI agent model developed by the Chinese AI company 阶跃星辰 (StepFun). It is distinguished as the industry's first GUI agent designed for deployment on mobile devices, enabling on-device execution of complex tasks. This is part of a larger family of models, including the cloud-based Step-GUI, which has also been recently upgraded.
| Model Name | Step-GUIEdge |
| Developer | 阶跃星辰 (StepFun) |
| Core Positioning | Industry-first open-source edge-side GUI agent for mobile deployment. |
The technical details of the Step-GUI model family are presented in the paper titled "Step-GUI Technical Report". The paper is available on arXiv.
- Paper Title: Step-GUI Technical Report
- arXiv URL: https://arxiv.org/abs/2512.15431
- Abstract:
Recent advances in multimodal large language models unlock unprecedented opportunities for GUI automation. However, a fundamental challenge remains: how to efficiently acquire high-quality training data while maintaining annotation reliability? We introduce a self-evolving training pipeline powered by the Calibrated Step Reward System, which converts model-generated trajectories into reliable training signals through trajectory-level calibration, achieving >90% annotation accuracy with 10-100x lower cost. Leveraging this pipeline, we introduce Step-GUI, a family of models (4B/8B) that achieves state-of-the-art GUI performance (8B: 80.2% AndroidWorld, 48.5% OSWorld, 62.6% ScreenShot-Pro) while maintaining robust general capabilities. As GUI agent capabilities improve, practical deployment demands standardized interfaces across heterogeneous devices while protecting user privacy. To this end, we propose GUI-MCP, the first Model Context Protocol for GUI automation with hierarchical architecture that combines low-level atomic operations and high-level task delegation to local specialist models, enabling high-privacy execution where sensitive data stays on-device. Finally, to assess whether agents can handle authentic everyday usage, we introduce AndroidDaily, a benchmark grounded in real-world mobile usage patterns with 3146 static actions and 235 end-to-end tasks across high-frequency daily scenarios (8B: static 89.91%, end-to-end 52.50%). Our work advances the development of practical GUI agents and demonstrates strong potential for real-world deployment in everyday digital interactions.
The open-source code for the Step-GUI project, named gelab-zero, is hosted on GitHub.
| GitHub URL | https://github.com/stepfun-ai/gelab-zero |
| GitHub Stars | 1.8k |
| License | MIT License |
The repository is structured into several key modules:
copilot_agent_client: Manages the client-side logic of the agent.copilot_agent_server: Contains the server-side logic.copilot_front_end: Provides the user interface components.mcp_server: Implements the Model Context Protocol (MCP) for standardized communication.tools: Includes various utilities and helper functions.examples: Offers practical examples and demonstrations of the agent's capabilities.
The Step-GUI models are available in two sizes: 4B and 8B parameters. The architecture is designed to be efficient and scalable, with a focus on real-world deployment.
- Calibrated Step Reward System: This self-evolving training pipeline is a major innovation, enabling the creation of high-quality training data from model-generated trajectories at a significantly lower cost.
- GUI-MCP (Model Context Protocol): A novel, hierarchical protocol for GUI automation that ensures user privacy by keeping sensitive data on-device. It combines low-level atomic operations with high-level task delegation.
- Training Data: The models are trained on a combination of existing benchmarks and a newly introduced benchmark, AndroidDaily, which is based on real-world mobile usage patterns.
The Step-GUI models have been benchmarked against several standard datasets, demonstrating state-of-the-art performance, particularly for the 8B model.
| Benchmark | Performance (8B Model) |
|---|---|
| AndroidWorld | 80.2% |
| OSWorld | 48.5% |
| ScreenShot-Pro | 62.6% |
| AndroidDaily (static) | 89.91% |
| AndroidDaily (end-to-end) | 52.50% |
The Step-GUI models are designed to be versatile and can be deployed across a range of platforms, including mobile, desktop, and automotive systems. The models can handle over 200 task scenarios, from simple queries to complex, multi-step operations.
- End-to-End Open Source: The project provides a complete, out-of-the-box solution, which significantly lowers the barrier for developers to enter the field of GUI agent development.
- Privacy-Centric Design: With its focus on local deployment and the innovative GUI-MCP, the model prioritizes user privacy.
- High Performance: The models have demonstrated leading performance on a variety of benchmarks, especially in realistic mobile environments.
- Cost-Effective Training: The Calibrated Step Reward System offers a breakthrough in reducing the cost and complexity of data annotation.
- Lack of Direct Comparisons: The available information does not provide direct, head-to-head comparisons with other specific GUI agent models, making it difficult to assess its relative performance in a competitive landscape.
- Potential for Bias: The performance data is primarily sourced from the developers' own technical report, which may introduce a degree of bias. Independent, third-party evaluations would be valuable for objective assessment.
[1] Step-GUI Technical Report. https://arxiv.org/abs/2512.15431 [2] stepfun-ai/gelab-zero on GitHub. https://github.com/stepfun-ai/gelab-zero