OpenMobile-2 shows open-source smartphone agents can match proprietary models
Researchers from Nanjing University, SenseTime and other universities released OpenMobile-2, an open-source smartphone agent built on Qwen3.5 (9 billion parameters) and Qwen3.6 (27 billion parameters). They trained the agents in a simulated environment containing more than 110 apps, including rebuilt versions of 35 popular Chinese services such as Taobao, JD, DiDi and Meituan, plus global tools…
Key points
- Training on over 110 simulated apps lifted unseen‑app success from 60.7 % to 67.5 %
- Larger OpenMobile‑2 model scored 59.6 points on SPA‑Bench, best among open models
- Direct commands improved task success by up to 10.3 percentage points and cut steps by about five
On real Android devices the larger model achieved 59.6 points on SPA‑Bench, the highest score reported for any open model, and scored 79.9 points on AndroidWorld after reinforcement learning. Direct‑command shortcuts, when explicitly trained, added up to 10.3 percentage points of task completion while reducing steps by about five. The team generated roughly 11,600 example workflows and over 2,000 practice tasks, using Google’s Gemini 3.1 Pro as the teacher model. The authors claim OpenMobile-2 ranks near the top among current smartphone agents, though they note the direct‑command feature remains experimental.
OpenMobile-2 proves open-source smartphone agents can rival proprietary models
The Decoder · 25 October 2026
Loading the full article…
This text was published by The Decoder and written by Jonathan Kemper. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Agents & Tools
All →- Agent Skills: Teaching AI 'How to Work' — What Changes with Anthropic's Donation? · 1 src
- Cognition releases open-source Devin Vulnerability Fixes plugin for PHP, Node.js, GitLab · 1 src
- OpenAI model spends testnet ethereum after locating a recovery phrase · 1 src
- Techcrunch lists AI agents that work over text messages · 1 src
- Simon Willison adds four‑player mode and Compass Rose map to Battledrones · 1 src
Comments
via GitHub Discussions