Technical Documentation: X-PLUG/MobileAgent
ℹ️ Provenance: Hybrid Fusion:
X-PLUG/MobileAgent(README + 1 In-Tree Chapters) · CodeWiki Reference · Recency: Active (< 180 days)
1. Project Overview & Quickstart (X-PLUG/MobileAgent)
👏 Welcome to try Mobile-Agent-v3.5 via our Modelscope online demo or Bailian online demo!
❗️We provide the Mobile-Agent-v3.5 API on Bailian for quick experience. View the documentation.
📢News
- [2026.5.12] 🔥🔥🔥 ToolCUA is released! An end-to-end Computer Use Agent for optimal GUI-Tool path orchestration, featuring a 2-stage training pipeline (trajectory-aware tool synthesis → Online Agentic RL) to master when to use GUI actions, invoke tools, and switch between them. See HuggingFace for the paper and model weights. Github.
[2026.3.31]🔥🔥 The Mobile-Agent-v3.5 is now available on Alibaba Cloud Wuying Cloud Phone (无影云手机) - a cloud-based Android environment for seamless Mobile Use experience. Learn More: Alibaba Cloud Wuying Cloud Phone | Documentation.[2026.3.19]🔥🔥 The GUI-Owl-1.5 series models are now available for online inference. Please refer to the Alibaba Cloud Bailian, and Modelscope API-Inference.[2026.2.14]🔥 GUI-Owl-1.5 is released, a new family of native multi-platform GUI agent foundation models (2B/4B/8B/32B/235B; Instruct & Thinking). The next-generation native GUI agent model family built on Qwen3-VL, supporting desktop/mobile/browser automation and achieving SOTA results on 20+ GUI benchmarks, with strong performance on end-to-end tasks, grounding, tool/MCP calling, and long-horizon memory. Model weights are available on HuggingFace. Technical report is avaliable on Link. See the GUI-Owl 1.5 README for details.[2025.11.25]The GUI-Owl series models are now available for online inference, thanks to Alibaba Cloud Bailian for providing computing power support. Please refer to the Link.[2025.10.30]We released OSWorld-MCP, which is a benchmark for evaluating Model Context Protocol (MCP) tool invocation capabilities in real-world scenarios. See the Link.[2025.9.24]We've released the demo on ModelScope that's based on Wuying Cloud Desktop and Phone. No need to deploy models locally or prepare devices, just input your instruction to experience Mobile-Agent-v3! ModelScope Demo Link and Bailian Demo Link. For a limited-time free Mobile-Agent-v3 API, please check the documentation. The new version based on Qwen-3-VL is coming soon.[2025.9.19]GUI-Critic-R1 has been accepted by The Thirty-ninth Annual Conference on Neural Information Processing Systems (NeurIPS 2025).[2025.9.16]We have released our latest work, UI-S1: Advancing GUI Automation via Semi-online Reinforcement Learning. The paper, code, dataset and model are now open-sourced.[2025.9.16]We've open-sourced the code of GUI-Owl and Mobile-Agent-v3 on OSWorld, AndroidWorld, and real-world mobile scenarios. See the OSWorld Code. The OSWorld RL-tuned checkpoint of GUI-Owl is also released. See the AndroidWorld Code and Real-world Scenarios Code.[2025.8.20]All new GUI-Owl and Mobile-Agent-v3 are released! Technical report can be found here. And model checkpoint will be released on GUI-Owl-7B and GUI-Owl-32B.- GUI-Owl is a multi-modal cross-platform GUI VLM with GUI perception, grounding, and end-to-end operation capabilities.
- Mobile-Agent-v3 is a cross-platform multi-agent framework based on GUI-Owl. It provides capabilities such as planning, progress management, reflection, and memory.
[2025.8.14]Mobile-Agent-v3 won the best demo award at the The 24rd China National Conference on Computational Linguistics (CCL 2025).[2025.3.17]PC-Agent has been accepted by the ICLR 2025 Workshop.[2024.9.26]Mobile-Agent-v2 has been accepted by The Thirty-eighth Annual Conference on Neural Information Processing Systems (NeurIPS 2024).[2024.7.29]Mobile-Agent won the best demo award at the The 23rd China National Conference on Computational Linguistics (CCL 2024).[2024.3.10]Mobile-Agent has been accepted by the ICLR 2024 Workshop.
📊Results
👀Features
📝Series of Work
- Mobile-Agent-v3.5 (Preprint): Multi-platform Fundamental GUI Agents. [Paper] [Code]
- Mobile-Agent-v3 (Preprint): Multi-modal and multi-platform GUI agent. [Paper] [Code]
- UI-S1 (ACL 2026 main): Advancing GUI Automation via Semi-online Reinforcement Learning. [Paper] [Code] [Dataset]
- GUI-Critic-R1 (NeurIPS 2025): A GUI-Critic for pre-operative error diagnosis method. [Paper] [Code]
- PC-Agent (ICLR 2025 Workshop): Multi-agent for multimodal PC operation. [Paper] [Code]
- Mobile-Agent-E (Preprint): Multi-agent for self-evolving mobile phone operation. [Paper] [Code]
- Mobile-Agent-v2 (NeurIPS 2024): Multi-agent for multimodal mobile phone operation. [Paper] [Code]
- Mobile-Agent-v1 (ICLR 2024 Workshop): Single-agent for multimodal mobile phone operation. [Paper] [Code]
📺Demo
text
<h3>Learn about Mobile-Agent-v3.5</h3>
<video src= "https://github.com/user-attachments/assets/97871c8c-5d1d-4f0f-aab0-84adddacbd0f"/>💻PC + 🌐Web
text
<h3>Search for the stock prices of Apple and Nvidia respectively. Then create a new spreadsheet in WPS Office. Enter the company names in column A and the retrieved stock prices in column B.</h3>
<video src= "https://github.com/user-attachments/assets/5dbc349b-0c99-4e66-a5c6-ad6d8accb933"/>
<h3>Create a new document in WPS Office and write a brief introduction about Alibaba with font size 12. Then search for Alibaba's logo in Edge browser, copy an image, and paste it at the end of the document.</h3>
<video src= "https://github.com/user-attachments/assets/8122baec-b5a1-49d8-a47d-cb613d21ec72"/>📱Phone
text
<h3>Today is Sunday, February 15, 2025. Search for flights from Guangzhou to Chengdu five days from now on Ctrip, check the cheapest flight, then search for the cheapest train ticker on the same route and tell me their prices.</h3>
<video src= "https://github.com/user-attachments/assets/4c5f29b1-c319-41bb-876d-9915bc3e215f"/>
<h3>Check the "魔搭ModelScope社区" (ModelScope Community) account on Xiaohongshu and Douyin, then tell me the total follower count across both platforms.</h3>
<video src= "https://github.com/user-attachments/assets/4ff3e951-f93a-49e8-b800-f060f765d67d"/>⭐Star History
📑Citation
If you find Mobile-Agent useful for your research and applications, please cite using this BibTeX:
text
/* Detailed source-code truncated for AI context efficiency. */2. In-Tree Documentation Chapters (X-PLUG/MobileAgent)
👏 Welcome to try Mobile-Agent-v3.5 via our Modelscope online demo or Bailian online demo!
❗️We provide the Mobile-Agent-v3.5 API on Bailian for quick experience. View the documentation.
📢News
- [2026.5.12] 🔥🔥🔥 ToolCUA is released! An end-to-end Computer Use Agent for optimal GUI-Tool path orchestration, featuring a 2-stage training pipeline (trajectory-aware tool synthesis → Online Agentic RL) to master when to use GUI actions, invoke tools, and switch between them. See HuggingFace for the paper and model weights. Github.
[2026.3.31]🔥🔥 The Mobile-Agent-v3.5 is now available on Alibaba Cloud Wuying Cloud Phone (无影云手机) - a cloud-based Android environment for seamless Mobile Use experience. Learn More: Alibaba Cloud Wuying Cloud Phone | Documentation.[2026.3.19]🔥🔥 The GUI-Owl-1.5 series models are now available for online inference. Please refer to the Alibaba Cloud Bailian, and Modelscope API-Inference.[2026.2.14]🔥 GUI-Owl-1.5 is released, a new family of native multi-platform GUI agent foundation models (2B/4B/8B/32B/235B; Instruct & Thinking). The next-generation native GUI agent model family built on Qwen3-VL, supporting desktop/mobile/browser automation and achieving SOTA results on 20+ GUI benchmarks, with strong performance on end-to-end tasks, grounding, tool/MCP calling, and long-horizon memory. Model weights are available on HuggingFace. Technical report is avaliable on Link. See the GUI-Owl 1.5 README for details.[2025.11.25]The GUI-Owl series models are now available for online inference, thanks to Alibaba Cloud Bailian for providing computing power support. Please refer to the Link.[2025.10.30]We released OSWorld-MCP, which is a benchmark for evaluating Model Context Protocol (MCP) tool invocation capabilities in real-world scenarios. See the Link.[2025.9.24]We've released the demo on ModelScope that's based on Wuying Cloud Desktop and Phone. No need to deploy models locally or prepare devices, just input your instruction to experience Mobile-Agent-v3! ModelScope Demo Link and Bailian Demo Link. For a limited-time free Mobile-Agent-v3 API, please check the documentation. The new version based on Qwen-3-VL is coming soon.[2025.9.19]GUI-Critic-R1 has been accepted by The Thirty-ninth Annual Conference on Neural Information Processing Systems (NeurIPS 2025).[2025.9.16]We have released our latest work, UI-S1: Advancing GUI Automation via Semi-online Reinforcement Learning. The paper, code, dataset and model are now open-sourced.[2025.9.16]We've open-sourced the code of GUI-Owl and Mobile-Agent-v3 on OSWorld, AndroidWorld, and real-world mobile scenarios. See the OSWorld Code. The OSWorld RL-tuned checkpoint of GUI-Owl is also released. See the AndroidWorld Code and Real-world Scenarios Code.[2025.8.20]All new GUI-Owl and Mobile-Agent-v3 are released! Technical report can be found here. And model checkpoint will be released on GUI-Owl-7B and GUI-Owl-32B.- GUI-Owl is a multi-modal cross-platform GUI VLM with GUI perception, grounding, and end-to-end operation capabilities.
- Mobile-Agent-v3 is a cross-platform multi-agent framework based on GUI-Owl. It provides capabilities such as planning, progress management, reflection, and memory.
[2025.8.14]Mobile-Agent-v3 won the best demo award at the The 24rd China National Conference on Computational Linguistics (CCL 2025).[2025.3.17]PC-Agent has been accepted by the ICLR 2025 Workshop.[2024.9.26]Mobile-Agent-v2 has been accepted by The Thirty-eighth Annual Conference on Neural Information Processing Systems (NeurIPS 2024).[2024.7.29]Mobile-Agent won the best demo award at the The 23rd China National Conference on Computational Linguistics (CCL 2024).[2024.3.10]Mobile-Agent has been accepted by the ICLR 2024 Workshop.
📊Results
👀Features
📝Series of Work
- Mobile-Agent-v3.5 (Preprint): Multi-platform Fundamental GUI Agents. [Paper] [Code]
- Mobile-Agent-v3 (Preprint): Multi-modal and multi-platform GUI agent. [Paper] [Code]
- UI-S1 (ACL 2026 main): Advancing GUI Automation via Semi-online Reinforcement Learning. [Paper] [Code] [Dataset]
- GUI-Critic-R1 (NeurIPS 2025): A GUI-Critic for pre-operative error diagnosis method. [Paper] [Code]
- PC-Agent (ICLR 2025 Workshop): Multi-agent for multimodal PC operation. [Paper] [Code]
- Mobile-Agent-E (Preprint): Multi-agent for self-evolving mobile phone operation. [Paper] [Code]
- Mobile-Agent-v2 (NeurIPS 2024): Multi-agent for multimodal mobile phone operation. [Paper] [Code]
- Mobile-Agent-v1 (ICLR 2024 Workshop): Single-agent for multimodal mobile phone operation. [Paper] [Code]
📺Demo
text
<h3>Learn about Mobile-Agent-v3.5</h3>
<video src= "https://github.com/user-attachments/assets/97871c8c-5d1d-4f0f-aab0-84adddacbd0f"/>💻PC + 🌐Web
text
<h3>Search for the stock prices of Apple and Nvidia respectively. Then create a new spreadsheet in WPS Office. Enter the company names in column A and the retrieved stock prices in column B.</h3>
<video src= "https://github.com/user-attachments/assets/5dbc349b-0c99-4e66-a5c6-ad6d8accb933"/>
<h3>Create a new document in WPS Office and write a brief introduction about Alibaba with font size 12. Then search for Alibaba's logo in Edge browser, copy an image, and paste it at the end of the document.</h3>
<video src= "https://github.com/user-attachments/assets/8122baec-b5a1-49d8-a47d-cb613d21ec72"/>📱Phone
text
<h3>Today is Sunday, February 15, 2025. Search for flights from Guangzhou to Chengdu five days from now on Ctrip, check the cheapest flight, then search for the cheapest train ticker on the same route and tell me their prices.</h3>
<video src= "https://github.com/user-attachments/assets/4c5f29b1-c319-41bb-876d-9915bc3e215f"/>
<h3>Check the "魔搭ModelScope社区" (ModelScope Community) account on Xiaohongshu and Douyin, then tell me the total follower count across both platforms.</h3>
<video src= "https://github.com/user-attachments/assets/4ff3e951-f93a-49e8-b800-f060f765d67d"/>⭐Star History
📑Citation
If you find Mobile-Agent useful for your research and applications, please cite using this BibTeX:
text
/* Detailed source-code truncated for AI context efficiency. */--- METRICS ---
- Files Extracted: 2
- Estimated Token Budget: ~5035 tokens
- Recency Window: Active (< 180 days)
- Canonical Reference: https://codewiki.google/github.com/X-PLUG/MobileAgent