OpenMobile-2: Open Phone Agents Hit 79.9 on AndroidWorld

Training an open phone agent on more apps did more for it than reinforcement learning did. OpenMobile-2’s 27B model scored 80.8 on AndroidWorld after supervised training and 79.9 after reinforcement learning. For teams building agents that operate apps, the release matters less for the score than for what it opens up: the training environments, data and recipe, the kind of material The Decoder notes has mostly been held by large companies.

What Happened

The Decoder reports that researchers from Nanjing University, SenseTime and other universities released two OpenMobile-2 models, built on Qwen3.5 (9B) and Qwen3.6 (27B). The OpenMobile-2 paper describes three parts:

• Environments: MobileGym++, with 35 commercial-style apps that support cross-app workflows, controllable resets and automatic verification. With Android emulator apps, the playground spans more than 110 apps.
• Data: nearly 12,000 interaction trajectories for supervised fine-tuning and more than 2,000 executable tasks with automatic rewards for reinforcement learning.
• App-native tools: more than 300 tools across 50+ apps, which agents can mix with ordinary screen actions.

The Decoder says the rebuilt apps are mostly China-focused, such as Taobao, JD, DiDi and Meituan, plus services like Slack and Google Drive. The training examples were generated by Google’s Gemini 3.1 Pro, so the open model still learned from a closed one.

Why It Matters

The benchmark gains are large. On AndroidWorld, the 27B model rose from 67.1 to 79.9 and the 9B model from 54.3 to 74, against 64.7 for the original OpenMobile. On MobileWorld, the 27B model went from 35.9 to 50.4. On SPA-Bench, which runs on real Android devices, it went from 31.9 to 59.6. The Decoder describes that as the highest score among open models, trailing only Gemini 3.1 Pro.

The more useful finding is about data. At the same data volume, training on more apps raised success on unseen apps from 60.7% to 67.5%, according to The Decoder. Reinforcement learning didn’t help much: the 27B model scored 79.9 on AndroidWorld after it, against 80.8 before.

Direct commands also mattered. The Decoder reports that models ignored app shortcuts like “add to cart” unless they were trained on them. Once trained, they solved up to 10.3 percentage points more tasks in about five fewer steps. The authors call this experimental, and it exists only in the simulated apps so far.

What This Means for Builders

App-operating agents have been hard to build in the open because the interaction data sat with large companies. Stack Archive covered a different open approach in OPPO’s on-device X-OmniClaw, which keeps sensing on the phone. OpenMobile-2 adds what X-OmniClaw didn’t: a reproducible training setup. Some results come from the team’s own 215-task benchmark, where a program checks app data instead of a language model judging the outcome. That avoids relying on a language model as judge, but it is still the authors’ own benchmark.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.