AutoGLM

What it is

AutoGLM is a voice-driven smartphone agent developed by Zhipu, the company also known as Z.ai^[1]. The description places it in a category of software that takes spoken input as its starting point and acts on a phone, rather than waiting for taps and typed commands. The point of such an agent is to shorten the distance between what a person says and what the device does: instead of navigating menus, the user states a goal and the agent is expected to carry it out on the phone.

Where it sits in Zhipu's line-up

AutoGLM is not a standalone experiment but one product among several that Zhipu offers. It sits alongside the company's GLM chatbot, its Ying video generator and its ZCode coding tool^[1]. Read together, that list describes a company covering conversation, video generation, programming assistance and phone-based task execution, with AutoGLM occupying the part of the portfolio that touches the handset directly.

Open releases

Alongside the product, Zhipu has put agent technology into public hands. On 8 December 2025 the company open-sourced AutoGLM-Phone-9B on Hugging Face, and followed it the next day with a multilingual version^[2]. The "9B" in the name follows the usual convention of naming a model after its parameter count, roughly nine billion weights.

AutoGLM is not the only agent work Zhipu has released this way. The company has also published research agents as open models, including WebRL in November 2024 and AndroidGen in May 2025^[2]. The public Hugging Face listing makes the release pattern visible in its checkpoints: it shows four WebRL checkpoints posted between 4 November and 9 December 2024, and two AndroidGen checkpoints posted on 28 May 2025^[2].

Why the open releases matter

Publishing weights and checkpoints changes who can examine a system. A phone agent is difficult to study from the outside, because most of the interesting behaviour happens in interaction with a device, an operating system and a user's spoken instructions. When the model behind it is downloadable, researchers and developers can inspect it, run it themselves and compare versions. The checkpoints listed for WebRL and AndroidGen also record intermediate stages rather than a single final model, which lets others see how the work developed over the period covered by the postings^[2].

For Zhipu, the combination is a familiar one in the current wave of agent development: a consumer-facing product on one side, and openly published models and research on the other. AutoGLM, WebRL and AndroidGen together show a company treating phone control and web-based task execution as related problems, and choosing to release much of the underlying work rather than keeping it entirely in-house^[2].