UI-TARS-1.5-7B

UI-TARS-1.5-7B

An open multimodal agent model that reads screenshots and decides what to do in apps, browsers, phones and games. For developers building computer-use and GUI automation agents.

Made by ByteDance-Seed on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: UI-TARS-1.5-7B is an open multimodal agent model from ByteDance-Seed, built on a vision-language model. It looks at a screen and works through tasks in virtual worlds, such as desktop apps, browsers, phones and games. It uses reinforcement learning to reason about a step before it acts.

Who it suits: It suits developers and researchers who build agents that operate a graphical interface from screenshots. The authors say it does well on computer, browser and phone tasks and on playing games.

Title: Who uses it
Industry | Who uses it | How they use it | Example
Software testing | QA engineers | Let an agent click through an app from screenshots | Run a check on a Windows program
Automation | Agent developers | Drive a desktop or browser with a vision model | Fill in a web form step by step
AI research | Researchers | Evaluate GUI agents on shared benchmarks | Compare results on OSworld and Android World

People who only need a text chatbot can skip it, since its strength is acting on screens.

Where to use it:

• Python with the transformers library
• Desktop, browser and phone environments the agent controls
• The UI-TARS-desktop application from the authors
• Windows, macOS or Linux, to download the files

Title: Top jobs
Use case | Who | What they get back
Computer use | Agent builders | Actions chosen from a desktop screenshot
Browser use | Developers | Steps for tasks on web pages
Phone use | Researchers | Actions for an Android-style screen
Game playing | Researchers | Moves in browser games

How to set it up:

• Python on your computer
• The huggingface_hub package
• About 33.2 GB of disk for the full weights, or about 4.1 GB at 4-bit by estimate
• An agent loop that sends screenshots and runs the model's actions

1. Install the Hugging Face hub package.
2. Download the model files to your computer.
3. Connect the model to a screen-control setup, such as the authors' desktop application.

pip install -U huggingface_hub
hf download ByteDance-Seed/UI-TARS-1.5-7B

How to use it: The card text here gives no run commands of its own. You send the model a screenshot and a task, and it reasons about the next step before it acts. For a ready-made front end, look at the authors' desktop application.

Title: Online benchmarks from the card
Benchmark | UI-TARS-1.5 | OpenAI CUA | Claude 3.7
OSworld, 100 steps | 42.5 | 36.4 | 28
Windows Agent Arena, 50 steps | 42.1 | not listed | not listed
WebVoyager | 84.8 | 87 | 84.1
Online-Mind2web | 75.8 | 71 | 62.9
Android World | 64.2 | not listed | not listed

Example: You give the model a screenshot of a settings page and ask it to turn on a switch. It thinks through what it sees, then names the action to take. The card shows demo videos of this kind of work but no text sample.

For newcomers and experts: Think of it as a model that sees your screen and suggests the next click or key press. A first try is to download the files, then run a single screenshot through it with a simple task.

For experts, the authors credit reinforcement learning for its step-by-step reasoning, which helps most at inference-time scaling. On grounding tests the card lists 94.2 on ScreenSpot-V2 and 61.6 on ScreenSpotPro, ahead of the OpenAI CUA and Claude 3.7 numbers it shows.

Take a screenshot of the screen
Send it to the model with a task
Model reasons, then picks an action
Your agent loop runs the action

Good to know:

• Agent model: it proposes actions, so you must supply the loop that carries them out.
• Language: the record lists English only.
• License: Apache 2.0.
• Cost to run: none named, and an 8.3B model needs a capable GPU.
• Data it reaches: it sees the screenshots you send, and it downloads files from Hugging Face.
• Not for: leaving it unsupervised on accounts or data you cannot afford to lose.

Title: OSworld score at 100 steps
UI-TARS-1.5: 42.5 points
OpenAI CUA: 36.4 points
Claude 3.7: 28 points

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 8.3B
• Weights: about 33.2 GB in F32; about 4.1 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 230,540
• Likes: 622
• License: apache-2.0
• Languages: en
• Quantized versions: 13
• Fine-tunes: 12
• Tags: transformers, safetensors, qwen2_5_vl, image-text-to-text, multimodal, conversational

Links

• Model: https://huggingface.co/ByteDance-Seed/UI-TARS-1.5-7B
• Files: https://huggingface.co/ByteDance-Seed/UI-TARS-1.5-7B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download ByteDance-Seed/UI-TARS-1.5-7B

Then follow the usage in its model card. Nothing in this kit runs by itself.