Xiaomi just published the technical details behind GUI-0, a mobile AI agent it claims works better in the real world because it was trained there.
Most GUI agents — software that controls a phone on your behalf — are built on vision-language models and tested in simulated environments or offline datasets. Xiaomi argues those proxies lie. Permission dialogs, payment screens, and account-state quirks that appear constantly in real apps barely show up in standard benchmarks, so models that ace tests often stumble in practice. GUI-0 addresses this with what the company calls a real-device-dominant hybrid infrastructure: physical phones are the primary training and evaluation environment, with sandboxes playing a supporting role only. The model goes through three training stages — supervised fine-tuning, step-level reinforcement learning, and a final agentic reinforcement learning pass — and feeds on an error-driven flywheel that converts failed attempts into corrected demonstrations and reflective explanations.
The benchmark numbers are respectable: 72.0% success on Xiaomi's in-house RealMobile suite and 78.9% on AndroidWorld, a public standard. More interesting is the claim of improved "execution stability" and abnormal-state recognition — the ability to recover when an app does something unexpected. That is the unglamorous work that separates a demo from a product.
Xiaomi is not alone here. Google, DeepMind, and a crop of startups are all racing to make phone-controlling agents that hold up outside a lab. The differentiator Xiaomi is betting on is the training loop itself, not a bigger model — which is either a genuine architectural insight or a story that sounds better before someone runs a rigorous independent comparison.