OpenMobile-2 Beats Gemini on AndroidWorld

A team from Nanjing University, SenseTime and four other universities has released OpenMobile-2, an open smartphone agent whose 27-billion-parameter model scored 79.9 on the AndroidWorld benchmark, ahead of Google's Gemini 3.1 Pro Preview at 70.7, according to the researchers' paper. The work, first posted to arXiv on Oct. 9, drew wider attention after The Decoder reported on it Oct. 25.

Key takeaways

• OpenMobile-2-27B scored 79.9 on AndroidWorld, up from 67.1 for its Qwen3.6-27B base model and above Gemini 3.1 Pro Preview's 70.7.
• The agent still trails closed models on harder tests: 50.4 on MobileWorld versus 70.1 for OpenAI's GPT-5.6 Sol, and 59.6 on real-device SPA-Bench versus 70.2 for Gemini.
• The team says it has opened the full recipe: a playground of more than 110 apps, 11,600 training trajectories, over 2,000 verifiable tasks and two model checkpoints.

What is OpenMobile-2?

OpenMobile-2 is a family of two agents that read a phone screen and tap, type and swipe to finish tasks. The 9B version is fine-tuned from Alibaba's Qwen3.5-9B and the 27B version from Qwen3.6-27B. Both were trained with supervised examples and then reinforcement learning.

The authors call it a "near-frontier mobile agent with fully open training environments and recipes." The training examples themselves came from a closed model: Gemini 3.1 Pro Preview wrote the task instructions and acted as the teacher agent.

A 110-app training playground

The core of the release is MobileGym++, a set of 35 newly built imitation apps modeled on commercial services, combined with 62 apps run in an Android emulator. The paper groups the apps into areas such as lifestyle, productivity, shopping, travel and finance.

App variety paid off in the team's tests. With a fixed budget of 2,000 trajectories, raising app coverage from 15 to 53 apps lifted the held-out AndroidWorld score from 60.7 to 67.5.

OpenMobile-2 benchmark results

Asterisks mark scores the authors reproduced themselves. None of the figures has been independently verified.

Apps that expose tools, not just screens

The paper also tests a hybrid setup in which apps offer selected functions as callable tools alongside their regular interface, with more than 300 tools across 50-plus apps. With tools enabled after hybrid training, the 27B model's success rate rose 10.3 points and its runs were 5.8 steps shorter on average.

To measure that, the team built MobileGym++ Bench, 215 long tasks checked by inspecting app data afterward. Of those, 106 require moving between apps.

Context

Phone agents are a crowded race, with model families such as MAI-UI and GUI-Owl among the open contenders the paper compares. The result also extends a year in which Chinese open-weight bases such as Qwen have pressed U.S. labs, as when Qwen3.8 Max climbed above every American lab but two on one agentic index.

What's next

The weights are on Hugging Face, but the paper and project page do not state a license for them, and the project page lists the original MobileGym data it builds on under a noncommercial CC BY-NC 4.0 license. Developers will want those terms clarified before building products on it.

FAQ

Can I download OpenMobile-2?

Yes. The 9B and 27B checkpoints are listed on Hugging Face, with code and the MobileGym++ environment on GitHub. A license has not been stated.

Is OpenMobile-2 better than Gemini or GPT on phones?

Only on AndroidWorld, by the authors' numbers. On MobileWorld and real-device SPA-Bench, Gemini 3.1 Pro Preview and GPT-5.6 Sol score higher.

References

• OpenMobile-2: Building Versatile Mobile Agents with Scalable Environments and App-Native Tools — arXiv (Oct. 9, 2026)
• OpenMobile-2 project page — OpenMobile-2 team
• OpenMobile-2 proves open-source smartphone agents can rival proprietary models — The Decoder (Oct. 25, 2026)