OpenMobile-2: Building Versatile Mobile Agents with Scalable Environments and App-Native Tools

Chenyang Yan; Yingying Zhang; Kanzhi Cheng; Qiushi Sun; Zhengyuan Pan; Zheng Ma; Hang Yan; Yian Wang

cs.SEOct 9, 2026

OpenMobile-2: Building Versatile Mobile Agents with Scalable Environments and App-Native Tools

Authors: Chenyang Yan, Yingying Zhang, Kanzhi Cheng, Qiushi Sun, Zhengyuan Pan, Zheng Ma, Hang Yan, Yian Wang,

Organizations: Nanjing University · SenseTime · The University of Hong Kong · Xi’an Jiaotong University ·

Abstract

We introduce OpenMobile-2, a near-frontier mobile agent with fully open training environments and recipes. We make three key advances: (1) Diverse environments with simulated commercial apps: We build MobileGym++, featuring 35 realistic, functionally rich commercial-style apps with cross-app workflows, while preserving full controllability for reset and verification. Together with newly configured apps in Android emulators, this yields a diverse playground spanning over 110 apps. (2) Open training data at scale: Building on this foundation, we curate nearly 12K mobile interaction trajectories for supervised fine-tuning and the largest open collection of verifiable RL training data for mobile agents, comprising over 2K executable tasks with automatic rewards. (3) Hybrid GUI and app-native tool use: We explore an experimental mobile-use setting where apps expose selected functionalities as app-native tools alongside their GUIs, allowing agents to interleave GUI actions and tool calls within a task. We implement this setting in MobileGym++ with over 300 carefully scoped tools across 50+ apps. Additionally, we introduce MobileGym++ Bench for complex, long-horizon commercial mobile scenarios, supporting both GUI-only and hybrid GUI-tool evaluation on a shared task suite. OpenMobile-2 performs competitively across established benchmarks, including AndroidWorld (79.9) and MobileWorld (50.4), while showing promising transfer to real-device mobile use, nearly doubling SPA-Bench performance from 31.9 to 59.6. All resources are available at https://os-copilot.github.io/OpenMobile2-Home/.

Figures & tables

Figure 1: OpenMobile-2 performs strongly across established mobile agent benchmarks and achieves leading results among open-source models on real-device evaluation.

Panel of Figure 2 — Figure 2: (a) Application coverage of OpenMobile-2. Each square represents two applications; dark and light tiles denote simulator and emulator apps, and dots mark apps with native tools. (b) Hybrid GUI and app-native tool use. The same task can be completed through GUI-only interaction (top) or by interleaving GUI actions with app-native tool calls (bottom) over a shared app state.

Environment	Platform	Scale		Interaction			Full Controllability
Environment	Platform	# Comm. Apps	# Tasks	Cross-app	GUI	App-native Tools	Full Controllability
AppWorld ( Trivedi et al., 2024 )	API	9	750	✓	✗	✗	✓
AndroidLab ( Xu et al., 2025 )	Emulator	5	138	✗	✓	✗	✗
AndroidWorld ( Rawles et al., 2025 )	Emulator	0	116	✓	✓	✗	✗
SPA-Bench ( Chen et al., 2025 )	Real-device	58	340	✓	✓	✗	✗
MobileWorld ( Kong et al., 2026 )	Emulator	4	201	✓	✓	✗	✗

Table 1: Comparison of mobile agent environments across scale, interaction, and controllability. ✓ denotes supported, ✗ unsupported, and

\sim

indicates limited system level cross-app interaction support in MobileGym. # Comm. Apps and # Tasks denote the number of covered commercial apps and benchmark tasks, respectively; Detailed application statistics are provided in Appendix C . MobileGym++ provides diverse, high-fidelity commercial-style applications with full controllability, while introducing an experimental app-native tool interface.

Figure 3: App-level interaction richness. Each marker denotes one app, positioned by its interactive actions and page transitions; marker size reflects the number of Bench tasks using that app. Our new apps reach 1.48

\times

more actions and 1.91

\times

more transitions.

Figure 4: Diversity and coverage of OpenMobile-Data-v2. (a) Domain distribution across seven categories for emulator and simulator data. Gini coefficients quantify domain balance. (b) Pairwise cosine similarity among task goals. (c) Single-app and cross-app workflow composition.

Commercial & Closed-Weight Models
Method	AndroidWorld		MobileWorld		MobileGym		MobileGym++		SPA-Bench
	Pass@1 $\uparrow$	Pass@3 $\uparrow$	GUI	MCP	SR	PR	GUI	Hybrid	L3
GPT-5.6-Sol	–	–	70.1	–	–	–	–	–	68.1*
Gemini-3.1-pro-preview	70.7	–	58.1	–	–	–	–	–	70.2*
MAI-UI-32B	73.3	–	36.2	30.0	–	–	–	–	–
MAI-UI-235B-A22B	76.7	–	39.7	37.5	–	–	–	–	–

Table 2: Main results on established mobile agent benchmarks and real-device evaluation.

*

denotes results reproduced by us, and

\pm

denotes the standard deviation across evaluation seeds.

Panel of Figure 5 — Figure 5: Environment coverage and app-native tool use. (a) Broader coverage improves held-out AndroidWorld performance at a fixed 2K-trajectory budget. (b) App-native tools improve success and efficiency, with larger gains after Hybrid SFT.

Table 8

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Hyperparameter	SFT	RL
Framework	Megatron-SWIFT	verl
Objective	Next-token prediction	DAPO (group-relative)
Trainable parameters	LLM (vision frozen)	LoRA
Optimizer	Adam ( $\beta_{1}{=}0.9$ , $\beta_{2}{=}0.95$ )	AdamW
Weight decay / gradient clip	0.1 / 1.0	0.1 / 1.0
Learning rate	$5\times 10^{-6}$	$1\times 10^{-6}$

Appendix

Table 5: Training hyperparameters. SFT values are shared by both runs; – indicates that a setting does not apply.

	GUI-only run		Hybrid run
Source	Trajectories	Steps	Trajectories	Steps
Emulator	7,725	83,631	7,725	83,631
Simulator	2,144	22,857	3,918	34,951
Total	9,869	106,488	11,643	118,582

Appendix

Table 6: Trajectories in each SFT run. The hybrid run adds 1,774 simulator trajectories that interleave app-native tool calls, so the simulator count rises from 2,144 to 3,918.

Panel of Figure 6 — Figure 6: Per-goal maximum cosine similarity. (a) SFT goals against the matched benchmark. (b) RL task goals against MobileGym++ Bench. Vertical lines mark each mean and the 0.7 threshold.

Benchmark	Train goals	Mean	Max	Above 0.7
AndroidWorld	7,718	0.434	0.892	1.11% (86)
MobileWorld	7,718	0.391	0.851	0.39% (30)
MobileGym	3,916	0.458	0.826	0.46% (18)
MobileGym++ Bench	3,916	0.504	0.861	2.25% (88)

Appendix

Table 7: Maximum cosine similarity of each SFT training goal to the matched benchmark. Above 0.7 counts goals whose maximum exceeds 0.7.

Domain	MobileGym++	Emulator
Lifestyle	18	19
Productivity	15	23
Entertainment	9	15
Shopping	8	1
Travel	5	3
Social	6	0

Appendix

Table 8: Application counts by domain in the two runtimes.

Domain	Applications
Lifestyle (18)	Baicizhan ∗ , Cainiao ∗ , Dianping ∗ , Keep ∗ , Luckin Coffee ∗ , Maoyan ∗ , Meituan ∗ , Meiyou ∗ , Weather † , Answer Sheet † , Browser † , Calculator † , Calculator 2 † , Clock † , Compass † , Gallery † , Settings † , Themes †
Productivity (15)	BOSS Zhipin ∗ , Days Matter ∗ , TickTick ∗ , Fanqie ToDo ∗ , flomo ∗ , Google Drive ∗ , Mail ∗ , Slack ∗ , Daily ∗ , Tencent Meeting, Calendar † , Phone † , Files † , Notes † , Messages †
Entertainment (9)	Podcasts ∗ , Fanqie Novel ∗ , Tencent Video ∗ , NetEase Cloud Music † , Ximalaya ∗ , Youku ∗ , Bilibili, Spotify, WeRead
Shopping (8)	JD ∗ , Pinduoduo ∗ , Suning ∗ , Taobao ∗ , Taobao Instant Commerce ∗ , Vipshop ∗ , Xianyu ∗ , eBay
Travel (5)	Ctrip ∗ , DiDi ∗ , Qunar ∗ , Maps, Railway 12306
Social (6)	Douban ∗ , Weibo ∗ , RedNote, Reddit, WeChat, X

Appendix

Table 9: The 63 MobileGym++ clients. ∗ marks the 35 applications built in this work and † marks the 16 system clients.

Domain	Applications
Lifestyle (19)	Camera, Chrome, Clock, Settings, Simple Gallery Pro, AndroidWorld, Broccoli, Clipper, MiniWoB, OpenTracks, APKPure, Breezy Weather, Coursera, DuckDuckGo, F-Droid, NerdCalci, Snapseed, Quark, Xiachufang
Productivity (23)	Files, Contacts, Dialer, Simple SMS Messenger, Simple Calendar Pro, Joplin, Markor, Tasks, Simple Draw Pro, Audio Recorder, Amaze, Material Files, MiXplorer, MT Manager, Etar, Omni Notes FOSS, Code Editor, DeepL, GeminiAssist, Doubao, Element, Mattermost, Session
Entertainment (15)	VLC, Retro Music, iQIYI, Youku, Bilibili, NetEase Cloud Music, NewPipe, AntennaPod, Toutiao, Yahoo News, Yahoo Sports, Wikipedia, Librera, 2048, Angry Birds Friends
Shopping (1)	eBay
Travel (3)	OsmAnd, Booking.com, Citymapper
Finance (1)	Pro Expense

Appendix

Table 10: The 62 Android emulator applications, grouped by domain. AndroidWorld and MiniWoB denote the task-support apps installed with the AndroidWorld suite.

Figure 8: MobileGym++ Bench composition. Left: tasks by the number of applications they touch. Right: task incidences by domain. Domain counts are not mutually exclusive.

Domain	Task incidences
Productivity	97
Finance	34
Shopping	53
Social	35
Entertainment	26
Lifestyle	46

Appendix

Table 11: MobileGym++ Bench task incidences by domain. A task that touches more than one domain is counted in each, so the column does not sum to 215.

Mobile Agent Benchmarks

See all 19 papers

OpenMobile: Building Open Mobile Agents with Task and Trajectory Synthesis
Apr 16, 2026Kanzhi Cheng, Zehao Li, Zheng Ma +11Synthetic Data Generation Mobile GUI Agents
1Nanjing University · 4Shanghai AI Laboratory · 2SenseTime +3
MobileGym: A Verifiable and Highly Parallel Simulation Platform for Mobile GUI Agent Research
May 25, 2026Dingbang Wu, Rui Hao, Haiyang Wang +8Mobile GUI Automation GUI Agents
Institute of Automation, Chinese Academy of Sciences · Peking University · The Chinese University of Hong Kong
MAI-Gym: An Open All-in-One Platform for GUI Agent Evaluation and RL Training across Mobile and Computer Environments
Oct 14, 2026Panrong Tong, Quyu Kong, Pengxiang Li +12RL for GUI Agents Mobile GUI Agents
Alibaba Token Hub, Alibaba Group

Apr 16, 2026J/K move · Enter open · S save

OpenMobile: Building Open Mobile Agents with Task and Trajectory Synthesis

Kanzhi Cheng, Zehao Li, Zheng Ma, Nuo Chen+10

1Nanjing University · 4Shanghai AI Laboratory · 2SenseTime+3

Synthetic Data Generation Mobile GUI Agents LLM Agent Trajectory Synthesis Multimodal Agents

Open paper AskPDF arXiv

Mobile agents powered by vision-language models have demonstrated impressive capabilities in automating mobile tasks, with recent leading models achieving a marked performance leap, e.g., nearly 70% success on AndroidWorld. However, these systems keep their training data closed and remain opaque about their task and trajectory synthesis recipes. We present OpenMobile, an open-source framework that synthesizes high-quality task instructions and agent trajectories, with two key components: (1) The first is a scalable task synthesis pipeline that constructs a global environment memory from exploration, then leverages it to generate diverse and grounded instructions. and (2) a policy-switching strategy for trajectory rollout. By alternating between learner and expert models, it captures essential error-recovery data often missing in standard imitation learning. Agents trained on our data achieve competitive results across three dynamic mobile agent benchmarks: notably, our fine-tuned Qwen2.5-VL and Qwen3-VL reach 51.7% and 64.7% on AndroidWorld, far surpassing existing open-data approaches. Furthermore, we conduct transparent analyses on the overlap between our synthetic instructions and benchmark test sets, and verify that performance gains stem from broad functionality coverage rather than benchmark overfitting. We release data and code at https://njucckevin.github.io/openmobile/ to bridge the data gap and facilitate broader mobile agent research.

Abstract

Mobile Agent Benchmarks

Explore similar work

OpenMobile: Building Open Mobile Agents with Task and Trajectory Synthesis

MobileGym: A Verifiable and Highly Parallel Simulation Platform for Mobile GUI Agent Research

MAI-Gym: An Open All-in-One Platform for GUI Agent Evaluation and RL Training across Mobile and Computer Environments