Editorial illustration for OpenMobile-2 AI Agents Rival Proprietary Models in Open-Source Release
Open-Source AI Agents Match Proprietary Models
Researchers from Nanjing University, SenseTime, and a handful of partner universities just put a number on a problem that's dogged open-source AI for years: smartphone agents built outside big tech labs keep falling behind because the training data they need sits locked inside companies like Google and Apple. Their answer is OpenMobile-2, a project that hands out two full agent models, built on Qwen3.5 at 9 billion parameters and Qwen3.6 at 27 billion, plus the entire training environment, dataset, and setup instructions needed to rebuild it from scratch.
The bigger finding isn't about model size. It's about variety. The team trained their agents inside a simulated environment stocked with more than 110 apps, and when they tested what actually moves performance on apps the model has never seen, raw data volume wasn't it.
Running the same amount of training data through a wider spread of apps did the work instead. That detail reshapes how agent training gets approached going forward, and it's why the researchers spent so much effort rebuilding dozens of apps rather than just scraping more logs from the ones they already had.
Training environment diversity matters more than sheer data volume here. The researchers held the amount of data constant and only increased the number of apps the agent practiced on, which pushed the success rate on unseen apps from 60.7 to 67.5 percent.
Why this matters
For anyone building mobile agents without a proprietary model budget, OpenMobile-2 is the clearest evidence yet that the bottleneck was never compute or parameter count, it was app diversity in training. A jump from 60.7 to 67.5 percent on unseen apps just from adding more simulated environments, holding data volume constant, is the kind of finding that should change how teams allocate engineering time: building 110 app environments is tedious work, but it beats scraping more interaction logs from the same handful of apps. The fact that a 27-billion-parameter Qwen3.6 model hit 59.6 on SPA-Bench and reportedly leads current smartphone agents, open or closed, is a direct challenge to the assumption that agent quality tracks model size or corporate backing.
We'd treat the "near the top" claim with some caution until independent teams reproduce it on their own hardware, but the authors releasing the full environment, data, and instructions removes most of the usual excuse for skepticism. Worth watching: whether other labs adopt the same app-diversity-first training recipe, and whether it generalizes beyond Android.
Related Reading
Open Source D&B rebuilds 642 million‑business database after AI agents hit limits
Open Source Box CEO Aaron Levie Says AI Will Drive "Biggest Cybersecurity Upgrade
Open Source Study: 6.7% of Deepfake Requests on Hugging Face Targeted Children
From Archives Ollama's New Phonon Feature Reads AI Replies Aloud
From Archives Cognition Releases Open-Source Agent to Patch Critical Node.js RCE Vulnerability
Common Questions Answered
What are the two OpenMobile-2 agent models and what parameters do they use?
OpenMobile-2 provides two full agent models: one built on Qwen3.5 at 9 billion parameters and another on Qwen3.6 at 27 billion parameters. Both models are open-source releases designed to rival proprietary smartphone agents developed by major tech companies.
How did training environment diversity impact the success rate of OpenMobile-2 agents on unseen apps?
Researchers demonstrated that training environment diversity matters more than data volume by holding the amount of data constant while increasing the number of apps the agent practiced on. This approach pushed the success rate on unseen apps from 60.7 to 67.5 percent, showing that app diversity is the critical factor.
What was the main bottleneck preventing open-source smartphone agents from matching proprietary models?
The primary bottleneck was not compute power or parameter count, but rather the lack of app diversity in training data. OpenMobile-2's findings reveal that access to diverse training environments across multiple apps is what enables open-source agents to achieve performance comparable to proprietary models from companies like Google and Apple.
Which institutions collaborated on the OpenMobile-2 project?
Researchers from Nanjing University, SenseTime, and several partner universities collaborated to create OpenMobile-2. This multi-institutional effort resulted in the release of complete agent models along with the entire training environment and datasets.
What does OpenMobile-2 release include besides the agent models?
OpenMobile-2 releases not only the two trained agent models but also the entire training environment and datasets used to develop them. This comprehensive release enables other researchers and developers to build and improve mobile agents without relying on proprietary data locked inside major tech companies.
Further Reading
- OpenMobile-2: Building Versatile Mobile Agents with Scalable Environments and App-Native Tools - arXiv
- OpenMobile: Building Open Mobile Agents with Task and Trajectory Synthesis - arXiv
- PalmClaw: A Native On-Device Agent Framework for Mobile Phones - arXiv
- Oppo's X-OmniClaw Brings Edge-Native AI Agents to Android, and the Architecture Choices Matter More Than the Demo Reels - NewClawTimes
