OpenMobile: Building Open Mobile Agents with Task and Trajectory Synthesis

The key detail the summary skips is the mechanism: OpenMobile doesn't rely on human-annotated demonstrations, which are expensive and hard to scale. Instead, it generates its own training data by having an agent explore a live Android environment and record what works, then uses a two-model policy where a weaker 'learner' model attempts tasks and a stronger 'expert' model steps in when the learner fails, creating a self-improving data loop.

The timing here is notable. The same week OpenAI expanded Codex with desktop automation and computer control (covered in 'OpenAI takes aim at Anthropic with beefed-up Codex'), and MM-WebAgent introduced hierarchical multimodal coordination for web tasks, OpenMobile is arriving from the research side with a fully open alternative for mobile environments. Where the Codex updates are closed, commercially positioned products, OpenMobile represents the academic track trying to close the capability gap through synthetic data rather than proprietary scale. These two tracks rarely merge quickly, but the benchmark proximity matters: 70% on AndroidWorld is close enough to commercial baselines that the gap is no longer dismissible.

Watch whether any of the major open-weight model teams (Mistral, Meta, or the Qwen group) adopt OpenMobile's trajectory synthesis pipeline within the next two quarters. Adoption there would signal the framework is genuinely reusable rather than a one-lab result tied to specific model assumptions.