AutoGLM-Phone-Multilingual

// Overview

AutoGLM-Phone-Multilingual (officially the AutoGLM-Phone-9B-Multilingual weights) is the international-facing variant of Z.ai's open-source mobile agent line, released alongside the Open-AutoGLM project in December 2025. It is a ~9B-parameter vision-language model fine-tuned from GLM-4.1V-9B-Base, with an architecture identical to GLM-4.1V-9B-Thinking. The model looks at a phone screenshot, reasons about what to do next, and emits a concrete UI action.

Unlike a chat model, AutoGLM-Phone-Multilingual is the brain inside a closed loop: the Open-AutoGLM framework captures the Android screen, sends it to the model, and the model returns one of a fixed action set (Launch, Tap, Type, Swipe, Back, Home, Long Press, Double Tap, Wait, plus a hand-back-to-human request for logins or captchas). Those actions are executed on a real device or emulator over ADB (Android Debug Bridge). The 'Multilingual' weight is tuned for English and other-language apps such as Gmail, Google Maps, Amazon, eBay, Booking.com, X, TikTok and WhatsApp, while the sibling AutoGLM-Phone-9B targets 50+ high-frequency Chinese apps like WeChat, Taobao, Douyin and Meituan.

Z.ai (Zhipu AI) ships the weights under the MIT license and the surrounding framework code under Apache-2.0, explicitly so enterprises and developers can self-host and keep screen data, logs and permissions inside their own environment. You can download the model from Hugging Face and ModelScope, run it locally with vLLM or SGLang, or call it through hosted APIs (Z.ai, ModelScope, Novita).

// Benchmarks

1. AndroidWorld (success rate, MobileRL-9B, the RL method behind AutoGLM-Phone)80.2%
2. AndroidLab (success rate, MobileRL-9B, the RL method behind AutoGLM-Phone)53.6%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

Pricing shown is Novita AI's hosted serverless rate for the AutoGLM-Phone-9B-Multilingual weights. Z.ai's own developer platform listed the model as 'free for a limited time' at launch. Weights are open and free to self-host (MIT license).

Pricing source ↗

// Strengths

• Fully open weights (MIT) plus open framework code (Apache-2.0) — self-hostable, no vendor lock-in
• Small 9B footprint runs locally via vLLM/SGLang and on consumer hardware through quantized GGUF/MLX builds (Ollama, LM Studio, llama.cpp, Jan)
• Purpose-built for on-device phone control: reads screenshots and emits concrete tap/type/swipe actions over ADB
• Multilingual tuning targets English-language apps (Gmail, Google Maps, Amazon, eBay, Booking.com) alongside Chinese ones
• Privacy-by-design positioning: screen data and permissions stay inside the user's own deployment
• Backed by the MobileRL online-RL training method, which reached state-of-the-art mobile-agent success rates

// Best for

• Autonomous mobile task completion (search a product, order food, book a flight) from a single natural-language instruction
• Self-hosted phone-use agents where screen data must stay on-premise for privacy or compliance
• QA / UI test automation across real Android apps and emulators
• Research on GUI agents and reinforcement learning for mobile control
• Building accessibility or hands-free assistants that operate everyday apps on the user's behalf

// How to access