I. What exactly have we been trying to achieve?
For a long time, we have been obsessed with answering one single question:
If AI is truly an "Assistant," can it actually pick up a phone like a human and finish a task from start to finish?
In our vision, AI shouldn't just live inside a chat box. It should step out and walk into the Apps you actually use every day:
- It should help you navigate a food delivery app, from selection to checkout;
- It should manage your cloud phone, batch-processing notifications, likes, and comments;
- It should handle the sales, customer service, and attendance workflows—automatically finishing those repetitive actions you dread doing yourself.
This is what AutoGLM is about: Teaching AI the true art of "Device Agency."
II. 32 Months: From Chaos to Control
To put it simply: We wanted AutoGLM to not just "speak," but to "act."
To make that sentence a reality, we started from scratch in April 2023—back when most people hadn't even heard of Large Language Models—and spent 32 months exploring every single detail.
1. From "Random Taps" to "Precision Control"
In our earliest versions, the system built on the large model only understood basic actions like "tap" or "swipe." It could occasionally finish a short workflow, but more often, it would get lost in nonsensical operations or fall into infinite loops.
To fix this, we spent nearly a year mapping out every possible failure. We tried to turn those "clumsy hands" into "capable hands":
- We built a comprehensive Phone Use capability framework;
- We abstracted the most fundamental actions: clicking, scrolling, typing, taking screenshots, and UI understanding;
- We taught the model to decompose a natural language command into a stable, replayable sequence of operations;
- We taught it to navigate the "environmental friction" of the real world: network jitters, intrusive pop-ups, and the chaos of ad overlays.
On October 25, 2024, we released the first AutoGLM capable of stably completing a full operation chain on a real device. It was regarded by the industry as the world's first AI Agent with true Phone Use capabilities.
2. The First "Digital Cash Gift" Sent by AI
In November 2024, AutoGLM achieved a milestone: it sent the first AI-automated "Red Packet" (digital cash gift) in human history.
It wasn't a script, and it wasn't an internal API call. It was the AI "seeing" the screen, "understanding" the context, and clicking through the banking interface step-by-step.
To us, this was a signal: From now on, many interactions on your phone can finally be fully handed over to AI.
3. Moving to the Cloud: A Safer Sandbox
In 2025, we released AutoGLM 2.0. We validated the scaling laws of Reinforcement Learning (RL), introducing MobileRL, ComputerRL, and AgentRL algorithms. This allowed AutoGLM to learn simultaneously across thousands of virtual environments, drastically expanding the Agent's accuracy and generalization capabilities.
More importantly, we didn't want the Agent to operate recklessly on a user's real phone or personal WeChat. So, we chose to place it in a virtual phone, detached from the user's physical reality:
- This phone runs in the cloud;
- Every action can be replayed, audited, and intervened upon;
- Truly sensitive data remains strictly isolated.
The intuition behind this design is simple: Before AI learns to use a phone, we must ensure it doesn't reach into places it shouldn't touch.
III. Why Open Source? What do we truly care about?
From a product perspective, AutoGLM is already powering many real-world scenarios. From an engineering perspective, our accumulated knowledge is enough to fill a stack of technical reports.
So, why choose to open source it at this specific moment? There are three things we have thought through clearly.
1. It is not enough for one company to do this alone.
The "AI Phone" is already a trend. But if the capability of "Phone Use" remains in the hands of a few manufacturers, it implies two things:
- The innovation space for developers will depend on whether platforms are willing to open their APIs;
- The most important device in a user's daily life will increasingly become a "Walled Garden"—a gateway controlled by others.
The primary intent of open-sourcing AutoGLM is to turn this capability into a public foundation that the entire industry can own and polish together.
You can treat it as a building block. Put it into your system, or tear it apart, rewrite it, and transform it. From that moment on, it is no longer just "Z.AI's AutoGLM"—it is part of what you and your team have built.
2. Returning Privacy and Control to the User
We are well aware that once Phone Use capabilities scale, they inevitably touch the most sensitive parts of the world: Personal chats, payments, logs, photo albums, and internal enterprise systems.
We do not want, nor should we have, these things in our hands.
Through open source and private deployment, enterprises and developers can fully control their data, logs, and permissions within their own compliant environments. It allows the phone to become an AI phone that exclusively belongs to you.
In one sentence: Technology is open to the ecosystem; Data and Privacy remain forever on the user's side.
3. Turning our path into the starting line for the "Agent Era"
AutoGLM has been a steep climb. We've chewed through technical problems, stepped into pitfalls, and rewritten frameworks. This path is reusable in many scenarios, and it deserves to be reused. But "a single flower does not make a spring." The explosion of Agents requires everyone to participate.
What we would love to see is:
- Teams building truly AI-Native phones based on AutoGLM;
- Researchers extracting a module to write a paper or create a new algorithm;
- Individual developers turning a demo into a passion project that runs perfectly in a niche scenario.
We hope that starting today, everyone can own their own mobile Agent.
IV. What can you get right now?
What we are open-sourcing is a suite of "ready-to-use" capabilities, not just a conceptual explanation. Specifically, this includes:
- Pre-trained Core Models & Inference Code
- The "Phone Use" Capability Framework & Toolchain
- Runnable Demos covering 50+ high-frequency Chinese Apps
- Android Adaptation Layer & Example Projects
- Documentation & Quick Start Guides
The models are released under the MIT License, and all code is under the Apache-2.0 License, hosted on GitHub.
You can use it as ready-made infrastructure, cannibalize parts of it, or modify it beyond recognition—as long as it helps you get closer to the "AI Native Phone" in your mind.
V. What's Next?
The current AutoGLM is not a perfect answer. It hasn't seen all the phones and apps in the real world yet. The future AI phone ecosystem will certainly take forms even more surprising than what we imagine today.
What we have done is simply handed a complete answer to the question "Can AI use a phone?"—honestly and authentically—into your hands.
At this moment, at the end of 2025, as Andrej Karpathy recently noted, we are facing not just the "Year of the Agent," but arguably the "Decade of the Agent."
Moving forward, AutoGLM will keep pushing. Let us drive Agent open source and research together. So that the "Jarvis" of our dreams will finally appear, for real, beside every one of us.