HomeAI models
ByteDance: UI-TARS 7B
Bytedance
•
UI-TARS 7B
•
Released 
April 2025

Bytedance
UI-TARS 7B

Open-weight 7B vision-language agent that reads screenshots and outputs GUI actions for browser, desktop, and mobile automation.

Modality:
Text
Image
model ID
bytedance/ui-tars-1.5-7b
Start for free
arrow black icon
View API Docs
arrow white icon

Output Speed *

N/A
tok/s

Intelligence Index *

N/A
/ 100

Context Window *

128000
tokens

Input price

0.6
Anytoken

Output price

1.2
Anytoken
A Compact Open-Weight Agent That Operates Interfaces From Screenshots UI-TARS 7B is ByteDance Seed's open-weight UI-TARS-1.5-7B, a vision-language agent that reads screenshots and decides the next GUI action. It is the smaller, publicly released member of the UI-TARS-1.5 line, built on a Qwen2.5-VL-based architecture under Apache 2.0. ByteDance reports 27.5 on OSWorld and 49.6 on ScreenSpot-Pro for this size. It suits computer-use, browser automation, and GUI-grounding workloads where a low-cost executor model runs under a separate planner. Access UI-TARS 7B through AnyAPI.ai

Performance

GUI Grounding Is Where the 7B Model Outperforms Its Size

In ByteDance's own comparison, UI-TARS-1.5-7B scores 49.6 on ScreenSpot-Pro, ahead of the earlier UI-TARS-72B-DPO at 38.1. It also scores 27.5 on OSWorld versus 24.6. Both are provider-reported results. The 7B remains well below the full UI-TARS-1.5 (61.6 and 42.5), so it is a strong grounding model rather than a top end-to-end computer-use agent. In production, expect good element localization, while long multi-step desktop tasks still need verification, retries, or a stronger planner.

Benchmarks

Third-Party Scores Land Below ByteDance's Reported Grounding Numbers

The BenchmarkList aggregator records UI-TARS 7B at 27.4% on OSWorld, rank 53 of 72 tracked models, and about 35.7% on ScreenSpot-Pro, rank 48 of 59. Hcompany's Holo1.5 report, from a competing model team, lists 39.0 on ScreenSpot-Pro for UI-TARS-1.5-7B, versus ByteDance's 49.6. Differences in evaluation setup are a possible cause. OSWorld agrees closely across sources, so use it as the anchor and test grounding on your own screens.

Output Speed

*
N/A
tok/s

Intelligence Index

*
N/A
/ 100

MMLU *

Broad world knowledge and problem-solving
0
%

GPQA *

PhD-level scientific reasoning across physics, biology, chemistry.
0
%

HLE *

Adherence to multi-step structured instructions.
0
%

LiveCodeBench *

Tool-calling reliability in long agentic loops.
0
%

Technical Specifications

What the model supports

UI-TARS 7B accepts text and image input and returns text only. API providers list a 128,000-token context window and a separate 2,048-token maximum output. The small output cap fits short action responses but rules out long reports or code generation. The 128K context allows screenshot history across a multi-step session, though each screenshot consumes tokens. Weights are released under Apache 2.0 on Hugging Face at roughly 8B parameters.
Verified Specifications — 
UI-TARS 7B
*
Input modalities
Text
Image
output modalities
Text
Context window
128000
 tokens
Maximum output tokens
2048
Reasoning
No
Knowledge cutoff
April 2025
Pricing (standard)
0.6
 AnyTokens in
 / 
1.2
 AnyTokens out

Limitations & Trade-offs

Where UI-TARS 7B falls short

1
Short output cap. Providers list a 2,048-token maximum output against a 128,000-token context window. That suits single GUI actions and brief reasoning text, but not long reports, extended code, or document generation. If a workflow needs long structured output, pair this model with a general-purpose LLM and use UI-TARS 7B only for screen interpretation and action selection.
2
Gap to the full UI-TARS-1.5. ByteDance reports 27.5 on OSWorld for the 7B versus 42.5 for the full UI-TARS-1.5, and 49.6 versus 61.6 on ScreenSpot-Pro. The model card also says the 7B is not specifically optimized for game scenarios. Long desktop workflows and game-like interfaces are therefore weaker fits than the headline UI-TARS-1.5 results suggest.
3
No verified API extras. The model card and provider listings document image-plus-text input and text output, but verified function calling or schema-constrained output support is not documented for this model. Treat responses as free-form action text that your application must parse and validate. Workflows that require guaranteed JSON schemas are better served by models that document structured outputs.
4
Modest speed and reproducibility. Aggregators report output speeds of roughly 19 to 35 tokens per second depending on route and snapshot, so sequential multi-step tasks accumulate latency. Third-party ScreenSpot-Pro results also fall below ByteDance's reported 49.6. Real-time or high-precision professional-software automation should be tested on your own screens first.

Best-Fit Workloads

Where this model earns its place

01

GUI element grounding

‍
The model's strongest evidence is screen localization: ByteDance reports 49.6 on ScreenSpot-Pro, ahead of UI-TARS-72B-DPO at 38.1. It fits pipelines that turn an instruction and a screenshot into a target element or coordinate. Third-party runs report lower grounding scores, so validate it on your own resolutions and applications before relying on it for dense professional interfaces.

02

Browser and desktop automation executor

‍
With 27.5 on OSWorld, the model can drive computer-use steps. It works best as a low-cost executor beneath a planner that decomposes the task and checks results. It suits RPA-style flows, form filling, and file operations in controlled environments. For long, open-ended desktop jobs, the full UI-TARS-1.5 or a stronger agent is the safer choice.

03

Screen content question answering

‍
Because it accepts screenshots, the model can answer questions about interface content. Hcompany's Holo1.5 report lists UI-TARS-1.5-7B at an 85.12 average across four screen-QA benchmarks, including VisualWebBench and WebSRC. That supports use in UI testing, accessibility checks, and screenshot triage, though the 2,048-token output cap limits long descriptions.

04

Open-weight agent research and prototyping

‍
The Apache 2.0 weights, around 8B parameters, make the model practical to evaluate, fine-tune, or self-host alongside an API deployment. Teams building GUI-agent prototypes can test prompts through an API first and move to local weights later. Verify license terms and your own accuracy requirements before shipping.

Pricing in anytokens via AnyAPI
Input
0.6
₳
Output
1.2
₳
Cache write
—
₳
Cache read
—
₳
Get API access
arrow black icon
Read Docs

on this page

PerformanceBenchmarksTechnical SpecificationsLimitations & Trade-offsBest-Fit WorkloadsIntegrationFAQ

Integration

Access UI-TARS 7B via AnyAPI.ai

Access UI-TARS 7B through AnyAPI.ai using a unified API built for multi-model AI applications. Integrate UI-TARS 7B without maintaining a separate provider-specific connection, and keep the flexibility to test, switch, or combine models as your application requirements evolve.

01

One API integration

Access UI-TARS 7B and other AI models through the same API workflow instead of maintaining separate integrations for every provider.

02

Easy model switching

Test UI-TARS 7B against alternative models or switch models as your performance, capability, or cost requirements change without rebuilding your application around another provider API.

03

Flexible for production

Use UI-TARS 7B from experimentation through production while keeping your AI stack flexible as workloads, traffic, and model requirements evolve.

04

Multi-model applications

Use UI-TARS 7B for the workloads where it performs best and combine it with other models for tasks that require different capabilities, performance, or efficiency.

Start free
arrow black icon
Read Docs

Frequently Asked Questions

Answers to common questions about integrating and using this AI model via AnyAPI.ai

UI-TARS 7B is ByteDance Seed's open-weight UI-TARS-1.5-7B, a multimodal vision-language agent for graphical interfaces. It takes screenshots and text instructions and produces text output describing GUI actions. ByteDance released it in April 2025 under the Apache 2.0 license as the smaller, public model of the UI-TARS-1.5 family.

API providers list UI-TARS 7B with a 128,000-token context window and a 2,048-token maximum output. Some listings show 131,072 tokens for context. These are separate limits: the context window holds your prompt and screenshots, while the output cap restricts each response to short text.

Yes, UI-TARS 7B accepts image input alongside text, which is how it reads screens. Its output is text only. It does not generate images, audio, or video. Provider listings and the Hugging Face model card classify it as an image-text-to-text model.

ByteDance reports 27.5 on OSWorld and 49.6 on ScreenSpot-Pro for UI-TARS-1.5-7B, compared with 24.6 and 38.1 for UI-TARS-72B-DPO. Independent figures are lower on grounding: BenchmarkList shows about 35.7% on ScreenSpot-Pro, and Holo1.5's report lists 39.0. OSWorld is close at 27.4% on BenchmarkList.

UI-TARS-1.5-7B weights are published on Hugging Face under the Apache 2.0 license, which generally permits commercial use under its terms. The full UI-TARS-1.5 model was offered separately as early research access, so the open weights and the top-scoring model are not the same release.

* Benchmark data source: Artificial Analysis artificialanalysis.ai