OpenMobile-2, an open-source AI smartphone-agent project developed by researchers from Nanjing University, SenseTime and other institutions, is demonstrating that openly released models can compete with proprietary systems at operating mobile applications. The project combines two AI agents, a training environment covering more than 110 apps, nearly 12,000 example workflows and more than 2,000 executable practice tasks to improve how AI systems interact with smartphones.
The researchers report that their larger 27-billion-parameter model scored 79.9 on AndroidWorld and 59.6 on SPA-Bench, a benchmark involving real Android devices. The results suggest that the diversity of applications an agent practices on can be just as important as the amount of training data it receives. By releasing models, training data, environments and training recipes, the team aims to make mobile AI automation more accessible to developers outside large proprietary AI companies.
Key Takeaways
- Open-source mobile AI: OpenMobile-2 releases trained agents, data, simulated apps and training recipes.
- Two model sizes: The project includes models based on Qwen3.5-9B and Qwen3.6-27B.
- More than 110 apps: Its training environment combines simulated commercial-style applications with Android emulator apps.
- Strong benchmark results: The larger model scored 79.9 on AndroidWorld and 59.6 on SPA-Bench.
- App diversity matters: With the training-data volume held constant, increasing the number of apps improved success on unseen applications from 60.7% to 67.5%.
- Experimental tool use: Agents can combine normal screen interactions with app-native commands, improving task completion after specialized training.
- Important limitation: Benchmark performance does not guarantee reliable operation across every real-world smartphone app or account.
What Is OpenMobile-2?
OpenMobile-2 is an open-source project designed to train AI agents that can operate smartphone applications much as a human user would. Instead of answering questions about an app, the agent receives an instruction and attempts to complete the task by interacting with the interface.
For example, a user might ask an agent to find a product, add it to a shopping cart, check a calendar or update information across several applications. The agent must understand the task, identify the relevant controls, perform the required actions and determine whether the intended result was achieved.
Building such systems is difficult because smartphone interfaces vary widely. Buttons move, screens change, authentication interrupts workflows and tasks can require several applications to work together.
Leading proprietary models have benefited from substantial internal training data and development infrastructure. Open researchers often lack access to comparable collections of real-world interaction examples, particularly for commercial apps.
OpenMobile-2 attempts to address that disadvantage by making the training environment itself more scalable and reproducible.
How the Project Trains Smartphone Agents
The project combines three main components: simulated applications, training data and models trained to use them.
1. A training environment with more than 110 apps
The researchers developed MobileGym++, a simulated environment containing 35 rebuilt commercial-style applications. These include services resembling shopping, transportation, food delivery and workplace software.
The simulated applications are combined with existing apps running in Android emulators, creating a training playground spanning more than 110 applications.
The environment is designed to make experiments repeatable. Unlike real apps, where an order, purchase or booking can create an irreversible change, simulated workflows can be reset to a known starting state.
The environment also preserves selected information between applications. This allows researchers to test tasks that require several steps across different services rather than measuring only simple, single-screen interactions.
2. Nearly 12,000 example workflows
The team generated approximately 11,600 demonstration trajectories, which show agents how to complete tasks through sequences of actions.
It also released more than 2,000 executable practice tasks with automated verification. These tasks provide a way to check whether the agent actually achieved the requested result rather than merely claiming that it did.
According to the project documentation, Google’s Gemini 3.1 Pro was used as a teacher model to generate demonstrations. OpenMobile-2 therefore relies partly on examples produced by a proprietary model while releasing the resulting training resources for broader research use.
3. Two open agent models
The project includes two model sizes built on Qwen model families:
| Model | Base model family | Main purpose |
|---|---|---|
| OpenMobile-2 9B | Qwen3.5-9B | A smaller option with lower resource requirements |
| OpenMobile-2 27B | Qwen3.6-27B | A larger model targeting stronger task performance |
The parameter count indicates the approximate scale of each model, not its guaranteed accuracy or operating speed. Actual deployment requirements depend on the hardware, quantization and inference setup.
Benchmark Results: How Well Does OpenMobile-2 Perform?
The researchers evaluated the models across established smartphone-agent benchmarks, simulated commercial-app tasks and real-device tests.
| Benchmark | OpenMobile-2 9B | OpenMobile-2 27B |
|---|---|---|
| AndroidWorld | 74.0 | 79.9 |
| MobileWorld | Not specified here | 50.4 |
| SPA-Bench, real-device result | Not specified here | 59.6 |
Source: OpenMobile-2 project materials and The Decoder’s October 25, 2026 report. Results use benchmark-specific scoring scales and should not be compared as if they were percentages.
On AndroidWorld, the larger model achieved a score of 79.9. On MobileWorld, which evaluates longer tasks involving multiple applications, it reached 50.4.
The result that attracted particular attention was SPA-Bench, where the larger model scored 59.6 on real Android devices. The researchers describe this as the highest score among open models in their comparison.
The predecessor OpenMobile-1 scored 31.9 on the same SPA-Bench comparison, according to the project report. The improvement suggests that training environment diversity and better task preparation can help open agents transfer their skills beyond simulated applications.
However, these are reported benchmark results, not proof that the agent can reliably operate every smartphone or outperform every proprietary system in every task.
Why Training on More Apps Can Beat Simply Adding More Data
One of the project’s most notable findings concerns the variety of applications used during training.
The researchers held the amount of training data constant and increased the number of apps represented in that data. The agent’s success rate on previously unseen applications rose from 60.7% to 67.5%.
That is an improvement of 6.8 percentage points.
The finding suggests that agents benefit from experiencing different interface layouts, interaction patterns and task types. Repeating thousands of similar actions in a small set of apps may teach an agent to memorize familiar screens without learning how to generalize to new ones.
A broader training environment exposes the model to more variation. An agent may learn that a search field, confirmation button or navigation menu can appear in different forms while serving similar purposes.
For developers, this points to an important lesson: improving AI agents may require better coverage of the environments where they will operate, rather than simply increasing the volume of examples.
A Hybrid Approach: Screen Actions Plus Direct App Tools
OpenMobile-2 also explores a hybrid approach that combines graphical user interface actions with app-native tools.
A conventional smartphone agent may need to tap through menus to add a product to a cart. In the experimental hybrid environment, an app can expose a controlled function that performs the same operation directly.
The agent can then choose between navigating the interface and using the available tool.
The project includes more than 300 app-native tools across over 50 applications. The tools and the graphical interface operate on the same underlying application state, allowing researchers to compare both approaches on equivalent tasks.
The results show that tool access alone does not automatically improve performance. Agents initially tended to ignore shortcuts they had not been trained to use. After hybrid training, however, the system completed up to 10.3 percentage points more tasks while requiring approximately five fewer steps.
This feature remains experimental, and the simulated tools are not equivalent to unrestricted control over every real smartphone application.
What OpenMobile-2 Means for Developers and Businesses
OpenMobile-2 could lower the barrier to experimenting with mobile automation.
Researchers can examine the training process, reproduce experiments and adapt the models to new tasks. Developers may be able to build assistants that navigate apps, collect information and complete routine workflows without depending entirely on proprietary agent platforms.
Potential applications include:
- Automating repetitive tasks across several mobile apps.
- Testing mobile applications with reproducible workflows.
- Building research prototypes for personal smartphone assistants.
- Studying how agents transfer skills to unfamiliar interfaces.
- Evaluating combinations of screen interaction and structured tools.
Businesses will still need to consider reliability, latency, hardware requirements and data protection before using these agents in production.
A model that performs well in a benchmark may still struggle with unexpected pop-ups, frequent interface updates, account verification or unfamiliar workflows. Actions involving payments, messages, bookings or sensitive information should have explicit permissions and appropriate confirmation safeguards.
The Bigger Picture
OpenMobile-2 reflects a broader change in the open-source AI ecosystem. Competition is increasingly about more than releasing model weights: it also depends on the quality of training data, the environments used for learning and the methods used to verify results.
By releasing an integrated training stack, the researchers provide other teams with a foundation for developing and evaluating smartphone agents. Their findings also suggest that carefully designed training environments can help narrow the performance gap between open and proprietary systems.
The challenge now is transferring these results into everyday use. Real applications change frequently, and agents must handle unpredictable interfaces, privacy requirements and consequential actions. If open projects can continue improving reliability and generalization, smartphone automation may become accessible to a much wider range of developers.
FAQs
1. What is OpenMobile-2?
OpenMobile-2 is an open-source smartphone-agent project that releases AI models, training environments, demonstration data and training recipes for agents that operate mobile applications.
2. How many applications are included in its training environment?
The combined environment spans more than 110 apps, including 35 rebuilt commercial-style applications in MobileGym++ and additional applications running in Android emulators.
3. Does OpenMobile-2 outperform proprietary AI models?
The researchers report competitive results across several benchmarks, including a 79.9 AndroidWorld score and a 59.6 SPA-Bench result for the larger model. These results do not establish that it outperforms proprietary models in every scenario.
4. Is OpenMobile-2 completely open source?
The project releases models, data, environments and training recipes. However, individual components can have different licences. Developers should review the licence for each model, dataset and repository before commercial use.
5. Can OpenMobile-2 safely control a real smartphone?
It has been evaluated on real Android devices, but benchmark performance is not a guarantee of safe operation. Deployments should restrict permissions and require confirmation for sensitive actions such as payments, purchases and sending messages.
Looking Ahead
The next test for OpenMobile-2 is whether its benchmark gains translate into dependable performance across a wider range of real applications. Continued testing on unfamiliar apps, longer workflows and changing interfaces will help determine how robustly the models generalize beyond their training environments.
For the wider AI industry, the project reinforces the importance of open infrastructure and reproducible evaluation. Better training environments could help open-source smartphone agents become more capable without requiring every research team to build a complete training system from scratch.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
