AutoGLM Phone 9B Multilingual

Open-AutoGLM, released December 8, 2025 by Z.ai (Zhipu AI), pairs the AutoGLM-Phone-9B model with its Multilingual variant in a complete automation framework spanning 32 months of development. The agent captures live phone screenshots, reasons about UI state, and executes tap, swipe, type, and launch actions via ADB, HDC, or WebDriverAgent. The system supports 50+ Android and 60+ HarmonyOS apps plus experimental iOS. Safety features include a sensitive-operation confirmation step and manual takeover for logins or CAPTCHAs, with remote Wi-Fi debugging eliminating USB tethering. Weights ship under MIT and code under Apache-2.0 at the zai-org/Open-AutoGLM GitHub repository. The project emphasizes real-device agency over brittle scripting, positioning itself as a practical open alternative to proprietary phone agents for developers and enterprises.

AutoGLM-Phone-9B-Multilingual is the international-facing variant of Z.ai's open-source mobile agent line, released December 2025 as a 9B-parameter vision-language model fine-tuned from GLM-4.1V-9B-Base. It serves as the perception-and-action core of the Open-AutoGLM framework, capturing Android screenshots via ADB and emitting a fixed action set including Launch, Tap, Type, Swipe, Back, Home, Long Press, Double Tap, Wait, and a hand-back-to-human request. The Multilingual weights target English and other-language apps like Gmail, Maps, Amazon, eBay, X, TikTok, and WhatsApp. Z.ai ships the model weights under the MIT license and framework code under Apache-2.0, enabling self-hosting. The model is available via Hugging Face, ModelScope, and hosted APIs including Novita AI, with local serving through vLLM or SGLang. Reported benchmark scores include 80.2% on AndroidWorld and 53.6% on AndroidLab using the MobileRL-9B training method, with a 25K context window and 3K max output tokens.

AutoGLM-Phone-9B-Multilingual is an open-source 9B vision-language model from zai-org that serves as the perception core of the Phone Agent framework for automating Android smartphone tasks. The model is finetuned from zai-org/GLM-4.1V-9B-Base and shares its architecture with GLM-4.1V-9B-Thinking, using BF16 Safetensors under an MIT license for research and educational use. Phone Agent interprets smartphone screens via multimodal perception and executes actions through ADB, planning sequences from natural-language commands like "Open Xiaohongshu and search for food recommendations." The system adds sensitive-action confirmation, human-in-the-loop fallback for logins or verification codes, and remote ADB debugging over WiFi or network, with deployment instructions in the GLM-V repository.

Open-AutoGLM is Z.ai's mobile agent framework letting users describe tasks in natural language while a vision-language model interprets the live phone screen and plans next actions. Control happens through ADB for Android, HDC for HarmonyOS, and WebDriverAgent for iOS, with a sensitive-operation confirmation step and manual takeover support for logins and verification codes. The framework ships under Apache-2.0 and states it is for research and learning only. The underlying AutoGLM-Phone-9B model targets Chinese apps while AutoGLM-Phone-9B-Multilingual covers English and other languages. Developers can self-host either variant using vLLM or SGLang, or point the framework at hosted endpoints including Z.ai, Novita AI, or Parasail. The framework integrates with the Midscene.js UI-automation SDK and supports interactive or one-shot command-line task modes.

Zhipu AI's AutoGLM suite pairs the mobile-focused AutoGLM with AutoClaw for desktop systems, both designed to act autonomously across interfaces rather than merely converse. AutoGLM launched March 31, 2025 at the Zhongguancun Forum, built on GLM-4.5 and GLM-4.5V with end-to-end asynchronous reinforcement learning. AutoGLM 2.0 followed August 18, 2025, described as the world's first AI agent with true Phone Use capability. Each user receives a dedicated cloud phone and cloud computer enabling 24/7 asynchronous operation without occupying the physical device. In December 2025, Zhipu open-sourced the core AutoGLM-Phone model family, expanding accessibility beyond the closed cloud offering. The mobile agent uses pure visual screen understanding to decide taps, scrolls, and inputs, contrasting sharply with traditional RPA or GUI scripting approaches.