OpenMobile-2 proves open-source smartphone agents can rival proprietary models

Researchers from Nanjing University, SenseTime and several other universities have released OpenMobile-2, an open-source project for training AI agents that operate smartphone apps themselves. The release includes two agent models, a 9-billion-parameter model built on Qwen3.5 and a 27-billion-parameter model built on Qwen3.6, plus the training environment, the data and instructions for reproducing the setup. The stated goal is to narrow the gap between open and proprietary smartphone agents, which the researchers attribute largely to training data being concentrated inside large companies.
The centerpiece is a simulated training ground covering more than 110 applications. Thirty-five of these are rebuilt versions of popular apps, including Chinese services such as Taobao, JD, DiDi and Meituan alongside globally used ones like Slack and Google Drive. The rest come from existing simulations and apps running in an Android emulator. The team rebuilt apps rather than training on the real ones because logins, actual purchases and the inability to reset app state make real-app training expensive and hard to repeat. Human reviewers compared each rebuilt app against the original and kept refining it until workflows and interfaces matched.
The simulations are persistent in a limited sense: orders, budgets and travel plans survive across sessions, and data can move between apps through the clipboard. Each simulation also resets to its starting state automatically. The agent sees only screenshots and has no access to internal app data, which mirrors how a person uses a phone. From this setup the team generated roughly 11,600 example workflows and more than 2,000 practice tasks with automated success checks. Gemini 3.1 Pro acted as the teacher model producing the examples, meaning the open system is effectively distilled from a closed one.
Results on established benchmarks show substantial gains. On AndroidWorld, the larger model rises from 67.1 to 79.9 points, and the smaller one from 54.3 to 74. The predecessor OpenMobile-1 scored 64.7 on the same test. On MobileWorld, which covers longer tasks spanning multiple apps, the 27B model improves from 35.9 to 50.4. The largest jump is on real Android devices: on SPA-Bench the larger model climbs from 31.9 to 59.6 points, which the authors describe as the best result among open models and second only to Gemini 3.1 Pro.
The most interesting finding concerns what drives improvement. Holding the amount of training data constant and only increasing the number of distinct apps pushed the success rate on unseen applications from 60.7 to 67.5 percent. In other words, app variety beat extra examples drawn from a handful of apps at the same data budget. For anyone building agent training pipelines, this suggests that breadth of environment coverage may be a cheaper lever than raw volume, though the result is measured on the team's own benchmarks.
Reinforcement learning added on top produced only limited gains, mostly in environments resembling the training setup. On AndroidWorld the larger model actually scored slightly worse after RL, ending at 79.9 rather than 80.8 points. That is a useful caution: post-training optimization appears to have overfit toward familiar conditions rather than generalizing further.
An experimental thread goes beyond tapping and swiping. More than 50 of the simulated apps expose direct commands such as "add to cart," letting the agent skip menu navigation, and the agent switches freely between screen actions and these tools depending on what an app offers. On its own this helps little, because models tend to ignore the shortcuts without targeted training. Once trained on them, agents solved up to 10.3 percentage points more tasks while needing about five fewer steps, and the researchers say the skill transfers to tools never seen in training. During the RL phase, runs that included at least one direct command received a bonus factor, though reward still depended on task completion, and penalties applied when the agent falsely reported success. These interfaces exist only inside the simulated apps for now, and the researchers call the approach experimental.
Some results come from a custom benchmark of 215 tasks, nearly half spanning multiple apps. Completion is judged not by a language model but by a program comparing an app's stored data before and after each run, which is a more deterministic check than LLM-as-judge scoring. The authors also checked for overlap between training and test tasks and report only minor similarities.
The work sits in a broader debate about how mobile agents should act. Microsoft Research classified API-based agents as more reliable and GUI-based agents as more flexible in a 2025 study, noting that many mobile apps offer no external interface at all, which is exactly the gap the hybrid tool approach tries to address. Oppo's open X-OmniClaw takes another route, running on real devices and replaying previously demonstrated paths in apps like Taobao or Meituan via deeplinks. Among closed systems, Google has built smartphone control into Gemini 3.5 Flash. In practice, a hybrid of GUI actions plus app-exposed tools looks like a plausible direction, since it retains the generality of screenshot-driven control while capturing some of the reliability of structured APIs.
Why it matters: Teams evaluating mobile agents now have an open baseline with published environments, data and reproduction steps, which makes results easier to check and build on rather than take on faith. The app-diversity finding, if it holds beyond these benchmarks, points to investing in broad simulated app coverage instead of collecting more examples from a few apps. The reliance on Gemini 3.1 Pro as teacher and the modest RL gains are worth weighing, and it remains an inference that these open models will match proprietary agents on the long tail of real production apps.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Smartphone agents don't get better by processing more data. They improve by training on as many different apps as possible. The researchers behind OpenMobile-2 built an open training environment with more than 110 applications, including 35 replicas of popular apps. On real Android devices, their model scores highest among open agents, the authors say. The article OpenMobile-2 proves open-source smartphone agents can rival proprietary models appeared first on The Decoder .