AutoGLM-Phone-9B
Z.ai (Zhipu AI) · China · 2025
An open vision-language model that reads a phone screen and taps it for you — driven over the Android debug bridge, tuned for Chinese apps, and released for research use only despite an MIT label.
AutoGLM-Phone-9B is the model behind Open-AutoGLM, Z.ai's framework for handing a smartphone over to software. The user types a sentence — the publisher's own example is "open Xiaohongshu and search for food recommendations" — and the system screenshots the phone, works out what is on screen, plans the next tap or swipe, and carries the task through to the end. The control path is deliberately ordinary rather than clever: the framework drives an Android handset over ADB, the standard developer debug bridge, and a HarmonyOS NEXT handset over its HDC equivalent, with a separate route documented for iPhones. Debugging permissions have to be switched on by hand, and on Android the phone additionally needs a helper keyboard installed, because ADB cannot type into a normal input method. The framework can also attach over WiFi rather than a cable. Two safety behaviours are built in: sensitive actions ask for confirmation, and logins or verification codes hand control back to the human. Architecturally the model is not a new design. It is a fine-tune of GLM-4.1V-9B-Base and, in the publisher's own words, architecturally identical to GLM-4.1V-9B-Thinking: forty text layers over a twenty-four-layer vision tower, a 65,536-token context window, and video tokens as well as image tokens in the vocabulary. Two published papers stand behind it — the 2024 AutoGLM work on GUI agents and the 2025 MobileRL work on online reinforcement learning for mobile agents. Two cautions belong in any deployment decision. First, the name understates the download: the weights occupy 20.6 GB in bfloat16, which is roughly 10.3 billion parameters once the vision encoder is counted, not nine. Hugging Face displays 934,400 parameters for this repository, but that figure comes from a wrong value written into the weight index by the publisher, and should be ignored. Second, the licence is stated twice and inconsistently: the repository metadata says MIT, while the project's own notice restricts the work to research and study and prohibits use for unauthorised data access or system interference. Anyone planning to run it in production should resolve that contradiction with the publisher first. The base variant is tuned for Chinese apps. A sibling release, AutoGLM-Phone-9B-Multilingual, covers English and other languages.
▸News
▸Videos
No videos yet.
▸Reviews
No reviews yet. Be the first!