A new mobile AI agent skips asking its expensive model for permission on every tap, trading a few accuracy points for big savings in speed and cost.
Jev-Mobile splits mobile GUI automation into two jobs: a vision-language model sets a local goal every so often, and a smaller, purpose-built model called Jev repeatedly picks the actual taps and swipes needed to reach it, working within the action space defined by the phone's accessibility tree so one expensive VLM call can cover many cheap ones. On the AndroidWorld benchmark, Jev-Mobile completed 79% of tasks successfully, edging out the SeeAct-V baseline's 78% but trailing a step-wise VLM approach that calls the big model at every step and hits 84%. Among tasks it completed, Jev-Mobile cut end-to-end execution time by 32.7% and model API cost by 73.4% compared with that step-wise baseline.
That cost gap is the real story. Running a VLM on every single screen interaction is expensive and slow, which is exactly why production GUI agents often feel sluggish or get limited to simple flows. Jev-Mobile's results suggest a lot of that per-step reasoning is redundant, and that a cheap, typed decision model can handle the mechanical parts just fine.
Five accuracy points is not nothing. Whether that trade is worth it depends on the task: booking a rideshare can tolerate an occasional miss, moving money cannot. For now, this looks like a solid efficiency trick for lower-stakes automation, not a wholesale replacement for step-by-step reasoning where every action counts.