Tool Suggestions
OpenMobile-2: Open Phone Agents That Hit 79.9% on AndroidWorld
The most useful thing about OpenMobile-2 is not its benchmark score. It is that the whole recipe is open. Researchers from Nanjing University, SenseTime and partner universities published the paper on arXiv on October 9, 2026, and released the models, the training data and the training environments — the kind of material that, until now, mostly stayed inside large companies. Coverage followed on October 25.
OpenMobile-2 is a phone agent: an AI that looks at a phone screen and carries out tasks by tapping and typing, the way a person would. It comes in two sizes, built on Alibaba's Qwen models: a 9-billion-parameter version on Qwen3.5 and a 27-billion-parameter version on Qwen3.6. The larger model is the headline act, reaching 79.9% on AndroidWorld, the standard benchmark for operating Android apps.
Three parts, all open
First, the environment. The team built MobileGym++, a training playground with 35 re-implemented commercial-style apps that support cross-app workflows, controllable resets and automatic verification of task outcomes. Combined with apps configured in the Android emulator, the playground spans more than 110 apps — the diversity the researchers credit most for the agent's ability to generalise to apps it has never seen.
Second, the data. On that playground they curated nearly 12,000 interaction trajectories for supervised fine-tuning and more than 2,000 executable tasks with automatic rewards for reinforcement learning — which they describe as the largest open collection of verifiable reinforcement-learning training data for mobile agents.
Third, a hybrid way of acting. Instead of only tapping the screen, the agent can mix ordinary screen actions with app-native tools: more than 300 tools across 50-plus apps, such as searching for a course inside a fitness app. The team also introduced MobileGym++ Bench, 215 long-horizon cross-app tasks with persistent user states, where GUI-only and hybrid agents can be compared on the same tasks.
The results, with the caveats
The 27B model scored 79.9 on AndroidWorld and 50.4 on MobileWorld, which tests longer multi-app tasks. On real devices, it nearly doubled SPA-Bench performance from 31.9 for the base Qwen3.6-27B model to 59.6 — the highest score among open models in the paper's comparison. The smaller 9B model reached 74.0 on AndroidWorld. One curious detail from coverage of the paper: the supervised-only 27B version scored slightly higher on AndroidWorld (80.8) than the reinforcement-learning version (79.9), a reminder that more training steps do not always mean better scores.
Three honest caveats belong with those numbers. The training examples were generated by Google's Gemini 3.1 Pro, so the open model still learned from a closed one. The rebuilt apps are mostly China-focused — Taobao, JD, DiDi and Meituan — alongside services like Slack and Google Drive, so results may not transfer evenly to every app store. And benchmarks are not proof of everyday reliability: results differ between the Android emulator and real devices, and a score on a test suite is not a promise that the agent will handle your phone correctly.
For researchers and teams building phone agents, that is exactly why the release matters: the training environments, data and recipe are available to build on, test against and improve. For everyone else, it is a sign of where phone AI is heading — from assistants that answer questions about your phone to agents that operate it, with the open versions now close enough to the commercial ones to matter.