Hugging Face Releases OpenMobile-2: Building Versatile Mobile Agents with Scalable Environments

Summary:Hugging Face has released OpenMobile-2, a frontier mobile agent with fully open training environments and tools. Key innovations include: (1) Diverse environments with simulated commercial apps (MobileGym++ featuring 110+ apps); (2) Large-scale open training data, including nearly 12K interaction trajectories and over 2K executable tasks with automatic rewards; (3) Hybrid GUI and app-native tool use, enabling agents to alternate between GUI actions and tool calls within tasks. OpenMobile-2 demon

---

Technical Mechanism Analysis

The core breakthroughs of OpenMobile-2 lie in its fully open training environments and scalable toolset for mobile agents:

1. Diverse Environment Construction

   • MobileGym++: Features 35 simulated commercial apps with cross-app workflows and supports reset and verification functionalities.
   • Android Emulator Integration: Newly configured apps bring the total number of environments to over 110, providing a rich interaction playground for agents.

2. Large-Scale Open Training Data

   • Supervised Fine-Tuning Data: Nearly 12K interaction trajectories offer abundant learning materials for agents.
   • Verifiable RL Training Data: Over 2K executable tasks with automatic rewards ensure the reliability and reproducibility of the training process.

3. Hybrid GUI and App-Native Tool Use

   • Tool Invocation Mechanism: Apps expose selected functionalities as native tools, allowing agents to alternate between GUI actions and tool calls within tasks.
   • MobileGym++ Bench: Supports evaluation in complex, long-horizon commercial mobile scenarios, covering both GUI-only and hybrid GUI-tool modes.

Engineering Trade-offs and Performance

OpenMobile-2 demonstrates competitive performance across multiple benchmarks. For instance, it scores 79.9 and 50.4 on AndroidWorld and MobileWorld, respectively, showcasing its strength in simulated environments. More importantly, its transfer performance to real devices is impressive, with the SPABench score nearly doubling from 31.9 to 59.6. This indicates that OpenMobile-2 not only excels in simulations but also adapts effectively to the complexities of the real world.

However, the hybrid GUI and tool invocation mechanism also introduces new challenges. For example, how to coordinate effectively between tool calls and GUI operations, and how to manage state in long-horizon tasks, are areas that require further optimization.

Developer Adoption and Deployment Recommendations

For developers, the open environments and toolset of OpenMobile-2 offer great flexibility. Here are some recommendations:

1. Customize Environments with MobileGym++: Developers can extend or customize the apps in MobileGym++ to suit different application scenarios.
2. Optimize Hybrid GUI and Tool Invocation Strategies: Design reasonable GUI and tool invocation strategies based on task requirements to improve the efficiency and accuracy of agents.
3. Continuous Monitoring and Evaluation: Utilize the open data resources of OpenMobile-2 to continuously monitor and evaluate agent performance and optimize based on feedback.

Conclusion

The release of OpenMobile-2 marks a significant advancement in mobile agents with open environments and toolset. Its excellent performance in both simulated and real environments, as well as its ability to handle complex tasks, opens new possibilities for AI applications in the mobile domain.

---

Source: Hugging Face Daily Papers (2026-10-09)

— END —

Tags: #Hugging Face Follow #Mobile Agents Follow #OpenMobile-2 Follow #Reinforcement Learning Follow #Multimodal Interaction Follow