Definition
Plain language
Xiaomi's phone-controlling AI agent, trained mostly on hundreds of real physical phones instead of simulators.
As stated in the literature
A ~30B (3B-active) sparse mixture-of-experts GUI agent trained via SFT then step-level and agentic reinforcement learning, using a hybrid real-device-and-emulator fleet and a failure-harvesting flywheel that mines recovery data from the agent's own mistakes.
Why it matters: It matters because training on real devices and mining its own mistakes for recovery data aims to make a phone-controlling agent more reliable in actual use.
For example, it can be asked to open a settings app and toggle an option, having learned such steps largely from practice on real physical phones.
Heard on the show
“That collapse is the whole subject of today's paper — a technical report out of Xiaomi called Xiaomi-GUI-0.”Episode 189 — Why Phone Agents Ace the Test and Crash on Your Actual Phone