OpenMobile-2 agents approach proprietary scores on phone benchmarks
Original title: OpenMobile-2 proves open-source smartphone agents can rival proprietary models
The 27B model rose from 31.9 to 59.6 on real-device SPA-Bench, which the authors call the best open score.
Researchers from Nanjing University, SenseTime, and other universities released OpenMobile-2, agents built on Qwen3.5 (9 billion parameters) and Qwen3.6 (27 billion), with training data, code, and a resettable environment covering more than 110 apps. Holding data volume fixed, more app variety raised success on unseen apps from 60.7 to 67.5 percent. Gemini 3.1 Pro generated the example workflows. Later reinforcement learning slightly cut the larger model's AndroidWorld score from 80.8 to 79.9.
Unconfirmed: Benchmark rankings and the best-open-model claim are reported by the authors.
- On AndroidWorld the 27B model moved from 67.1 to 79.9; OpenMobile-1 scored 64.7.
- On MobileWorld the larger model improved from 35.9 to 50.4.
- The setup produced about 11,600 workflows and more than 2,000 practice tasks.
- Trained tool shortcuts solved up to 10.3 points more tasks in about five fewer steps.
- Those direct commands exist only in simulated apps and are called experimental.
- Organizations
- Nanjing UniversitySenseTime
From the original article
4 min read at the-decoder.com
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.
By Jonathan Kemper · the-decoder.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.