OpenMobile-2 smartphone agents match proprietary rivals

Researchers have released OpenMobile-2, an open-source smartphone agent system that performs competitively with proprietary models by prioritizing app diversity over raw data volume. The project, led by teams from Nanjing University, AI company SenseTime, and several other institutions, demonstrates that training agents on a wide range of applications matters more than simply scaling up training examples.

What changed

The team unveiled OpenMobile-2 along with two trained models built on the Qwen architecture: a 9-billion-parameter version and a 27-billion-parameter version. They released the complete training environment, data, and instructions needed to reproduce the work. The project includes fully open training environments and recipes.

The core innovation is MobileGym++, a training environment containing over 110 simulated applications. The team rebuilt 35 popular apps, primarily used in China such as Taobao, JD, DiDi, and Meituan, alongside globally recognized services like Slack and Google Drive. These apps were reconstructed to match the original workflows and interfaces, verified by human reviewers. Combined with existing simulations running in Android emulators, the playground provides a realistic but controllable space for agent training.

The researchers generated approximately 11,600 example workflows and over 2,000 practice tasks with automated success checks. Google’s Gemini 3.1 Pro served as a teacher model that produced the initial examples. The training data includes both supervised fine-tuning trajectories and reinforcement learning tasks, creating what the researchers describe as the largest open collection of verifiable mobile agent training data.

The experimental approach allows over 300 app-native tools across 50+ simulated apps to coexist alongside graphical interfaces. This enables agents to use direct commands like “add to cart” instead of navigating menus. According to the paper, agents using these tools reduced task steps by approximately five while improving success rates by up to 10.3 percentage points. The researchers note this capability also transfers to tools the models never encountered during training.

Why it matters

OpenMobile-2 addresses a significant gap in open-source AI. Mobile agents trained by large companies have access to vast proprietary datasets of real app interactions, giving closed models an advantage over open alternatives. This project proves that well-designed diversity in training environments can narrow that gap without requiring access to proprietary data or real-world app interactions at scale.

The findings challenge assumptions about how to build better AI agents. Most practitioners assume that more training data leads to better performance. The researchers instead held the amount of training data constant and varied the number of apps in the training environment. This experiment showed the success rate on unseen applications jumped from 60.7 to 67.5 percent, a substantial gain from app diversity alone. This insight matters for professionals building agents: environment design may be as important as data volume.

On established benchmarks, both models showed notable improvements. The larger model achieved 79.9 points on AndroidWorld, compared to the predecessor OpenMobile-1’s score of 64.7. On the MobileWorld benchmark covering longer tasks across multiple apps, it scored 50.4. The most significant gain appeared on real Android devices, where the larger model reached 59.6 points on SPA-Bench, a jump from 31.9 points in the previous version and the highest score reported among open models.

The project’s release of complete training recipes allows professionals to inspect, modify, and extend the training process. This transparency represents a distinct advantage over proprietary systems like Google’s Gemini 3.5 Flash, which includes smartphone control but remains a closed system. Organizations can now adapt the environment for domain-specific apps rather than accepting whatever a closed vendor provides.

The research also highlights a trade-off between data volume and diversity. Reinforcement learning produced only limited gains in the study. On AndroidWorld, the larger model actually scored slightly lower after reinforcement learning, reaching 79.9 instead of 80.8 points. This suggests diminishing returns from additional training phases when the base environment is already well-designed.

What to test

Before adopting OpenMobile-2 or using it as a baseline, professionals should verify several claims:

• Generalization on real apps: The benchmark scores come from simulated or emulated environments. Real-world performance on apps not in the training set needs independent testing. The 59.6 score on real devices comes from a custom benchmark with 215 tasks, not from established third-party evaluation. Test whether learned behaviors transfer to your specific applications.

• Tool-use adoption rates: The research shows that agents ignore app-native tools without targeted training. Verify whether your intended use case benefits from tool integration or whether pure graphical interface interaction suffices, since the tool feature is marked experimental.

• Data reset and repeatability: The simulated apps reset to starting state automatically, enabling clean training loops. Real apps rarely offer this capability. Test whether the learned behaviors transfer when reset mechanisms are unavailable or when persistent state carries between sessions.

• Cross-app workflows: The training environment supports data passing between apps via clipboard and persistent state like orders and travel plans. Confirm that your target workflows match this capability level.

• Geographic app coverage: The training set emphasizes apps popular in China (Taobao, JD, Meituan, DiDi) alongside global services. If your use case focuses on region-specific applications, check whether the diversity extends to your target apps.

• Reproducibility and setup cost: The team released all training recipes openly. Measure the computational and engineering effort required to reproduce or adapt the environment for your specific apps before committing to deployment.

Common questions

What is a mobile agent?

A mobile agent is an AI system that views smartphone or tablet screenshots and operates apps by performing actions like taps, swipes, and text input, similar to how a human user would interact with a device. The agent sees only what appears on screen and has no access to internal app data, mimicking real-world constraints.

How does OpenMobile-2 differ from proprietary smartphone agents like Google’s?

OpenMobile-2 is fully open-source with publicly released training environments, data, and model weights. Google has built smartphone control directly into Gemini 3.5 Flash, but that remains a closed system. OpenMobile-2 allows researchers and professionals to inspect, modify, and extend the training process, whereas proprietary models are not customizable in this way.

Can OpenMobile-2 agents run on real phones?

Yes, the larger model achieved its 59.6 score on real Android devices using SPA-Bench, a real-phone benchmark. The training environment is simulated for efficiency and repeatability, but the agents transfer to actual hardware. However, this requires additional setup and integration work beyond running the trained model in the simulated environment.

The conclusion

OpenMobile-2 demonstrates that open-source smartphone agents no longer need to lag significantly behind proprietary models when trained thoughtfully on diverse simulated environments. The project’s central finding that app diversity matters more than data volume offers a practical path for professionals who want control over their agent systems without relying on closed platforms.

The real-device performance gain from 31.9 to 59.6 points is substantial but still leaves room for improvement compared to frontier commercial systems. Teams considering deployment should test the model on their specific app portfolios and workflows before committing to production use. The experimental tool-use feature shows promise but needs validation in your context.

For organizations balancing cost, customization, and performance, OpenMobile-2 provides a credible open alternative. The released training recipes enable practitioners to rebuild or extend the environment for domain-specific apps rather than accepting whatever a closed vendor provides. Watch for community adaptations that might narrow the gap further, particularly around handling real-world app complexity and geographic diversity.

Like other agentic systems, OpenMobile-2 will improve as feedback loops tighten. The next phase likely involves tighter integration with existing agent security frameworks and broader app ecosystem support. For now, the project credibly shows that open approaches offer transparency and control, even if closed systems retain a performance edge.