OpenMobile-2: open smartphone agents; app diversity over raw data volume
Nanjing University, SenseTime, and collaborators introduce OpenMobile-2: MobileGym++ rebuilds ~35 commercial-style apps into a 110+ app playground; ~11.6K demos and 2K+ verifiable RL tasks; agents on Qwen3.5-9B and Qwen3.6-27B. Reported scores include AndroidWorld 79.9, MobileWorld 50.4, and SPA-Bench 31.9→59.6 on real devices. At a fixed data budget, more apps beat more trajectories on the same apps. Teacher demos are attributed to Gemini 3.1 Pro; hybrid GUI+app-native tools are experimental.
OpenMobile-2 (arXiv:2610.14157; v1 listed Fri 9 Oct 2026; affiliations include Nanjing University, SenseTime, HKU, XJTU, and others) argues open mobile agents lack large, resettable, verifiable interaction environments. The authors release MobileGym++ with 35 functionally rich commercial-style simulated apps (Decoder cites Taobao, JD, DiDi, Meituan plus Slack and Google Drive) plus newly configured emulator apps for a playground of over 110 apps—persistent state, cross-app flows, clean reset, screenshot-only observation.
On data and models they report nearly 12K mobile trajectories for SFT (project home: 11.6K) and over 2K executable RL tasks with automatic rewards; agent backbones are Qwen3.5-9B and Qwen3.6-27B. THE DECODER states demonstration trajectories were produced by Google’s Gemini 3.1 Pro as the teacher. Headline benchmarks from the abstract/paper: AndroidWorld 79.9 for the larger model (Decoder: 67.1→79.9; smaller 54.3→74); MobileWorld 50.4 (Decoder: 35.9→50.4); real-device SPA-Bench 31.9→59.6, described as near the top among open models. Treat figures as author/report claims.
A central analysis claim: with data volume held fixed, increasing the number of apps seen in training lifts success on unseen apps more than adding more trajectories on a few apps (Decoder: 60.7%→67.5% on unseen apps). An experimental hybrid GUI + app-native tool setting exposes 300+ scoped tools across 50+ apps in MobileGym++; without targeted training models often ignore shortcuts, while after training Decoder cites up to +10.3 percentage points success and about five fewer steps. OpenMobile2/OpenMobile2-Code and the project home list release plans for data, weights, and the environment. THE DECODER covered the work on 2026-10-25.
Sources
- OpenMobile-2: Building Versatile Mobile Agents with Scalable Environments and App-Native Tools — arXiv
- OpenMobile-2 PDF — arXiv
- OpenMobile-2: Building Versatile Mobile Agents with Scalable Environments and App-Native Tools — OpenMobile-2 project home
- GitHub - OpenMobile2/OpenMobile2-Code — GitHub / OpenMobile2
- OpenMobile-2 proves open-source smartphone agents can rival proprietary models — THE DECODER