Vision-language-action (VLA) model
A vision-language-action (VLA) model takes an image and an instruction and returns an action, which moves a robot arm in robotics and completes a step in an app on a phone or computer. AGI, Inc. builds VLA models for screens and is optimizing them for Snapdragon platforms with Qualcomm Technologies.
Not to be confused with a vision-language model (VLM), which returns text about an image instead of an action.
Last updated
Related terms
Mobile use
Mobile use is an AI agent's ability to operate the apps on a phone or tablet and finish tasks for a person.
Action model
An action model is an AI model that returns the next action in an app or on a screen instead of a written answer.
OSWorld
OSWorld is a public benchmark that tests computer-use agents on 369 tasks on real computers, across desktop apps, files and websites.
Computer use
Computer use is an AI agent operating a computer's apps and websites through the screen, the way a person does.
Input | An image, such as a camera frame or a screenshot, plus an instruction |
|---|---|
Output | An action, one step at a time |
Origin | Robotics research |
Research results | AGI-0: 100% on AndroidWorld (116/116 tasks). AGI's agent: 76.26% on OSWorld (October 2025) |
Silicon | Snapdragon platforms, with Qualcomm Technologies since March 2026 |
Where does the term come from?
The term comes from robotics, where VLA models turn camera images and instructions into motor commands. A VLA model for screens does the same with an interface. It takes what is on the screen and returns the next input.
How is a VLA model different from computer use?
Computer use names the job, operating a computer's interface. VLA names the kind of model, one that takes vision and language as input and returns actions.
Approach | What it acts on | What it returns |
|---|---|---|
Robotics VLA models | The physical world, through a robot | Motor commands |
General-purpose models with computer use | Screens, usually through a cloud session | Clicks and keystrokes |
Screen VLA models, such as AGI-ACT | Phone and computer interfaces, on the device or hybrid | Steps in an app |
How are VLA models for screens measured?
Benchmarks give the model real tasks on real software and check the end state. AndroidWorld has 116 tasks in 20 real Android apps, and AGI-0 completed all 116, a 100% score on the official leaderboard. OSWorld has 369 tasks on real desktop computers, and AGI's agent scored 76.26% in October 2025, where people score 72.36%.
FAQ
Is a VLA model the same as a computer-use model?
The terms overlap. Computer use describes the job, operating a computer's interface. VLA describes the kind of model, one that takes images and instructions and returns actions. AGI's models fit both descriptions.
Does AGI, Inc. build its own VLA models?
Yes. AGI-ACT is AGI, Inc.'s own family of action models for screens. The company is optimizing them for Snapdragon platforms under a collaboration with Qualcomm Technologies announced in March 2026.
Where can I read AGI's research?
AGI, Inc. publishes benchmark results and research notes on its blog, including the OSWorld result from October 2025 and the AGI-RUN inference benchmarks from July 2026.
