Skip to content

UI-TARS

A native GUI agent model for end-to-end task automation

Visit official site

What UI-TARS is

The model is described as a native GUI agent from ByteDance that integrates perception, reasoning, grounding and memory into a single vision-language model. It targets end-to-end task automation without predefined workflows or manual rules.

What you can do with it

  • Use a unified action framework across desktop, mobile and web environments
  • Handle complex tasks through multi-step trajectory and reasoning training
  • Monitor the GUI in real time and respond to changes as they happen
  • Apply task decomposition and reflection for multi-step planning and error correction

Application scenarios

  • Detect and fix GUI issues in software testing
  • Complete document processing and data entry in office automation
  • Browse pages, fill forms and extract information in web automation

Model notes

  • Perception, reasoning, grounding and memory are integrated into one vision-language model
  • System 1 and System 2 reasoning combine intuitive reaction with advanced planning
  • Large-scale annotation and synthetic datasets support generalization and robustness

Getting started

  • Visit Hugging Face Inference Endpoints
  • Build the input instruction with the provided Prompt template for mobile or computer scenarios
  • Base64 encode a local screenshot and send it with the instruction to the model interface
  • Read the returned reasoning result with an action summary and specific operations

Pros & cons

✓ What we like

  • Perception, reasoning, grounding and memory sit in one vision-language model
  • End-to-end automation needs no predefined workflows or manual rules

! What to watch out for

  • The documented entry point is Hugging Face Inference Endpoints
  • Local screenshots must be Base64 encoded before being sent to the model interface

FAQ

What does the model integrate?

The material states perception, reasoning, grounding and memory are integrated into a single vision-language model.

How is an input prepared?

The documented steps use a Prompt template, a Base64 encoded screenshot and the model interface.

Last reviewed: 2026-11-15

More LLM API platform tools

View all →

How we review