OpenMobile-2 Shows How Diverse App Training Helps Open Models Rival Top AI Agents
Researchers from Nanjing University and SenseTime release OpenMobile-2, proving that app diversity allows open AI agents to compete with proprietary models on real devices.
A team of researchers from Nanjing University, AI firm SenseTime, and partner institutions has released OpenMobile-2, a pair of open-source models designed to bridge the performance gap between open and proprietary smartphone AI agents. Built upon Qwen3.5 with 9 billion parameters and Qwen3.6 with 27 billion parameters, the project provides complete access to the underlying training environment, dataset, and code instructions. Rather than focusing strictly on increasing total data volume, the researchers demonstrated that expanding the variety of applications an agent practices on is the primary driver for improving real-world task performance.
Rebuilding app ecosystems for reproducible training
Training smartphone agents directly on live applications presents significant practical obstacles. Requirements such as user logins, real financial transactions, and an inability to easily reset application states make live system training both expensive and difficult to replicate. To bypass these issues, the research team manually reconstructed 35 widely used applications—including Chinese services like Taobao, JD, DiDi, and Meituan, alongside global tools such as Slack and Google Drive. Combined with existing software simulations and Android emulator environments, the entire practice setup covers more than 110 distinct applications.
Human reviewers verified that the reconstructed interfaces and workflows matched the original software. These simulated environments maintain state persistence across sessions—storing data like clipboards, orders, and travel plans—while resetting automatically between practice runs. The agent operates strictly via screenshots without direct access to internal application backends, mimicking human visual interaction. Using Google's Gemini 3.1 Pro as a teacher model, the team generated roughly 11,600 example workflows and over 2,000 practice tasks with automated success checking.
Experimentation showed that application diversity during training matters far more than raw data volume. When keeping data quantity constant while increasing the range of practice apps, the agents' success rate on previously unseen applications climbed from 60.7 percent to 67.5 percent.
Benchmark gains and hybrid navigation
The OpenMobile-2 models recorded substantial gains across standard benchmarks. On AndroidWorld, the 27B model reached 79.9 points—up from 67.1 points—while the 9B model rose from 54.3 to 74 points, outperforming the previous OpenMobile-1 score of 64.7. On MobileWorld, which tests longer tasks spanning multiple apps, the 27B model progressed from 35.9 to 50.4 points. The most pronounced jump occurred on real physical Android devices evaluated via SPA-Bench, where the 27B model climbed from 31.9 to 59.6 points, setting the highest score among open models and trailing only Gemini 3.1 Pro. Subsequent reinforcement learning yielded minimal additional improvements.
The study also explored a hybrid interaction approach where agents combined visual user interface navigation (tapping and swiping) with direct function calls, such as an "add to cart" command integrated into more than 50 simulated apps. Once trained on these shortcuts, the agents solved up to 10.3 percentage points more tasks while cutting execution lengths by approximately five steps. This capability successfully transferred to unfamiliar tools encountered during testing.
What it means for developers
OpenMobile-2 offers developers a fully reproducible framework for building practical GUI and tool-using agents. By releasing the complete codebase, simulated application suite, and model weights, the researchers lower the barrier to studying mobile automation.
For developers building autonomous software agents, the project highlights that training across diverse application environments is significantly more effective than generating massive quantities of data from a limited set of applications. Combining visual user interface interaction with direct backend tool commands enables agents to complete complex multi-app tasks faster and with fewer errors. As open models like Qwen and closed models like Gemini 3.5 Flash continue to advance agentic capabilities, developers can try top AI models cheaply through one API at https://apixoai.online to evaluate different model architectures for their own automation workflows.
Ultimately, OpenMobile-2 proves that open-source models paired with varied, well-structured training environments can achieve benchmark performance that directly competes with proprietary systems on real smartphone devices.
Source: OpenMobile-2 proves open-source smartphone agents can rival proprietary models — The Decoder. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
