UI-TARS 72B on live screen streams

What UI-TARS 72B is good at

UI-TARS was trained specifically for computer-use: given a screenshot and a goal, it outputs a concrete next action, like a click coordinate, a key press, or a scroll, rather than a free-form description of the screen. That native action grounding is what separates it from a general vision-language model asked to describe a UI.

It handles multi-step GUI tasks across desktop and web interfaces, tracking state across a sequence of screenshots rather than reasoning about a single image in isolation. On published GUI agent benchmarks it leads open models built for the same task.

• Screenshot-to-action grounding for GUI agents
• Multi-step task tracking across a sequence of screen states
• State-of-the-art results among open GUI agent models

UI-TARS 72B and the Overshoot workflow

UI-TARS 72B is not currently in Overshoot's live model catalog; GET /v1beta/models returns the current list. The workflow a computer-use model slots into is the standard one: publish a screen share over WebRTC through LiveKit, then send a chat-completions request referencing the latest frame with an ovs:// URL. The model reads the current screen state and returns the next action as its response, which an agent loop can execute before requesting the next frame.

Responses from live models stream over SSE, and thread_id keeps prompt caching warm across the many repeated, similar requests a GUI agent loop generates as it steps through a task on the same stream.

UI-TARS versus general vision models for agent tasks

General vision-language models like Qwen3-VL or InternVL3 can describe a UI in detail but were not trained to emit direct actions, so an application built on them needs extra logic to turn a description into a click or keystroke. UI-TARS skips that translation step, which is why it is the more direct choice for building a computer-use agent rather than a screen-understanding assistant.

Frequently asked questions

Can UI-TARS 72B analyze live video?

UI-TARS is built to read screen states, and a live screen share is its natural input, but it is not currently in Overshoot's live model catalog. Live models read a WebRTC screen share through ovs:// frame references in a chat-completions request, streaming responses back over SSE. Check GET /v1beta/models for the current catalog.

Is UI-TARS 72B open source?

Yes. UI-TARS 72B is released under the Apache 2.0 license by ByteDance, so weights are downloadable and usable commercially, including for self-hosted agent stacks.

What makes UI-TARS different from a general vision model?

UI-TARS was trained to emit concrete actions, like clicks and key presses, directly from a screenshot, rather than just describing what it sees. That makes it purpose-built for GUI agents, where a general vision-language model would need extra logic to convert a description into an action.

What is UI-TARS 72B best used for?

UI-TARS fits computer-use agents that need to complete multi-step tasks on a desktop or web interface, like filling out a form or navigating a settings menu. It is not the right choice for general scene description or open-ended visual question answering.

Related models

• Qwen3-VL 32B
• GLM-4.5V
• InternVL3 78B
• Molmo 72B
• Holo 3.1 35B A3B
• Holo3 35B A3B