Alibaba trained a model on a rack of 100 real phones — it now finishes 92 out of 100 tasks on an actual handset
Alibaba's Tongyi MAI team released Qwen-UI-Agent on 20 August, a model built to operate screens rather than answer questions — phones, desktops, browsers, and a command line, in one system. Instead of training only in simulation, the team ran a live rig of more than 100 physical smartphones covering 150-plus apps, and built MobileWorld-Real, a benchmark of 400-plus tasks on real devices. Reported results: 92.2 percent on MobileWorld-Real, 97.5 percent on AndroidDaily, 79.5 percent on OSWorld-Verified and 73.6 percent on WebArena, ahead of Claude Opus 4.8 and GPT-5.6 on mobile and browser work — though Opus still leads on the harder OSWorld-v2, 54.8 to 40. The smaller MAI-UI models at 2B and 8B are on Hugging Face under Apache 2.0.
Why it mattersThe gap between a demo agent and a useful one has always been the sim-to-real gap: an agent that clicks correctly in a simulator gets lost on a real phone with a real popup. Training against a rack of actual handsets is an expensive, unglamorous way to close it, and the numbers on real devices are the ones worth watching. The small versions being open means the technique does not stay inside Alibaba.
Read in the app — free, in 9 languages
Related stories
A 27-billion-parameter model beat Opus 4.8 and GPT-5.5 at reproducing published papers2026-08-23An eight-billion-parameter model that reads pictures and draws them, at 4K, with an Apache licence — and the community had it running in two days
2026-08-22OpenAI has given away the engine under Codex, and its own numbers say the engine is worth more than the model
2026-08-22Give a model the reference list of an unpublished paper and ask what the paper says. The best ones get it 15 percent of the time.
2026-08-22Google's open models have been downloaded a billion times, and outsiders have built 100,000 versions of them
2026-08-22