OpenMobile-2 releases open smartphone agents trained across more than 110 apps

What happened

Researchers from Nanjing University, SenseTime, and other universities are releasing OpenMobile-2 smartphone agents built on Qwen3.5 and Qwen3.6, alongside their training environment, data, and reproduction instructions. The larger model improves from 31.9 to 59.6 points on real-device SPA-Bench, which the researchers call the best result among open models. Holding data volume constant, increasing training app diversity raises success on unseen apps from 60.7 to 67.5 percent; experimental direct-command interfaces also improve task completion after targeted training.

Read next

• More changes in today’s briefingMoonshot launches Kimi K3.1 Pro with weights promised November 4 and faster Highspeed API

The story so far

Researchers from Nanjing University, SenseTime, and other universities have released two OpenMobile-2 smartphone agent models, along with their training environment, data, and reproduction instructions. The models use Qwen3.5 with 9 billion parameters and Qwen3.6 with 27 billion parameters. The larger model’s score on the real-device SPA-Bench benchmark rose from 31.9 to 59.6 points, which the researchers describe as the best result among open models. The researchers also tested how training across more varied apps affects performance. With training data volume held constant, increasing app diversity raised task success on unseen apps from 60.7 to 67.5 percent. Experimental direct-command interfaces also improved task completion after targeted training.