一个通用基础 GUI 智能体,面向移动应用、Web 平台和桌面操作系统——在同一个闭环的感知–推理–行动智能体中,协同扩展环境、任务与反馈。
🌟 UI-Venus-2 有哪些新变化
与 UI-Venus-1.5 相比,我们引入了:
- 📱 扩展的多语言移动环境:大幅扩充的可执行移动应用池,覆盖中英文应用生态,并配合深度研究驱动的查询生成策略,将任务查询建立在真实应用功能之上——提升所生成指令的准确性、有效性和可执行性。
- 🖥️ 从零构建的计算机使用能力:通过计算机使用数据采集和任务专项训练,从零打造专用的桌面操作系统能力,将 UI-Venus 系列扩展至在同一个统一的端到端智能体中实现移动、Web 和操作系统交互。
- 🎯 基于关键点的验证:任务完成情况依据与任务相关的视觉关键点来判定,而非对最终屏幕进行粗略的整体观察,并通过多模型投票聚合异构评判者——减少单一评判者的偏差,使奖励信号对奖励攻击具有鲁棒性。
- 🔄 验证增强反思: 经过验证的反馈被作为反思监督蒸馏回训练中,使智能体能够区分部分进展与真正完成,避免因观察误读而导致的过早终止,并在长时程交互中恢复。
📈 UI-Venus-2 基准测试表现
UI-Venus-2 在 GUI 智能体基准测试上的表现。 每个面板将 UI-Venus-2-27B 和 UI-Venus-2-9B 与一些选定的强基线进行对比。我们倾向于选择在最接近的可用任务子集和步数预算上评估的独立端到端系统;来源报告的动作脚手架可能仍存在差异。MobileWorld 使用 117 个任务、50 步的纯 GUI 成功率,WebVoyager 使用刷新后的 595 任务划分,Odysseys 使用 200 个任务的平均评分标准得分,VenusBench-CAPTCHA 使用全部 219 个样本的 micro Pass@1,VenusBench-GD 使用英文指令的 micro 平均准确率。“*”表示由我们复现的结果。
📰 新闻
- [2026/09] 我们发布了 VenusBench-CAPTCHA,这是一个用于在多样化 CAPTCHA 任务上评估 GUI 智能体的综合基准。请见分支 VenusBench-CAPTCHA。
- [2026/08] 我们发布了 UI-Venus-2,这是一个 9B/27B 的通用基础 GUI 智能体,通过规模化多语言环境、关键点锚定验证和验证增强反思,统一了移动端、网页端和桌面端交互。
- [2026/02] 我们发布 UI-Venus-1.5,这是一个端到端 GUI 智能体,专为稳健的真实世界应用而设计。
- [2026/02] 我们发布 VenusBench-Mobile,这是一个面向移动端 GUI 智能体的高难度在线基准测试。请查看分支 VenusBench-Mobile。
- [2025/12] 我们发布 VenusBench-GD,这是一个全面的多平台 GUI 定位基准测试。请查看分支 VenusBench-GD。
- [2025/8] 我们发布 UI-Venus,这是我们 UI 智能体模型的第一个版本。
🧭 概览
✨ 演示
🛠 Venus 框架
我们提供两个用于在真实环境中运行智能体的框架:
| 框架 | 描述 | 文档 |
|---|---|---|
| 移动端框架 | 用于单任务执行、多设备批量执行、轨迹记录与反思的 Android/ADB 智能体框架。 | English · 中文 |
| 浏览器插件 | Chrome 侧边栏扩展,将 UI-Venus 连接到当前活动的浏览器标签页,并以交互方式执行浏览器任务。 | English · 中文 |
有关目录结构和入口点,请参阅 Venus Framework 概览。除非各自的依赖要求另有说明,下面的轻量级领域示例无需任一框架即可使用。
🚀 快速开始
安装
conda create -n ui-venus-2 python=3.11 -y
conda activate ui-venus-2
pip install -r requirements.txt
需要 Python 3.10 或更高版本。以下所有命令均在仓库根目录下执行。通过 MODEL_URL、MODEL_NAME 以及 API_KEY 或 MODEL_API_KEY 配置 OpenAI 兼容的模型服务;相同的值也可以在每个领域脚本的顶部进行编辑。当两个密钥变量都设置时,API_KEY 优先。将 MODEL_NAME 设置为你的 UI-Venus-2 9B 或 27B 模型的服务名称。
移动端
对随附的预录屏幕截图序列运行多轮推理。N_IMG 控制保留多少张最近的历史屏幕截图:
MODEL_URL=http://127.0.0.1:8000/v1 \
MODEL_NAME=UI-Venus-2 \
N_IMG=2 \
bash scripts/mobile.sh
本示例仅执行模型推理,不在设备上执行操作。如需真机 ADB 自动化,请使用上文的 Mobile Framework。
Computer
对预录的桌面屏幕截图序列运行多轮 Computer 推理:
MODEL_URL=http://127.0.0.1:8000/v1 \
MODEL_NAME=UI-Venus-2 \
N_IMG=2 \
bash scripts/computer.sh
默认命令使用随附的桌面屏幕截图示例。该独立示例会校验并规范化模型操作,但不会在主机上执行这些操作。它不依赖 OSWorld 运行。
Browser
按照领域文档所述,以 CDP 端口启动 Chrome,然后运行一个自然语言浏览器任务:
MODEL_URL=http://127.0.0.1:8000/v1 \
MODEL_NAME=UI-Venus-2 \
bash scripts/browser.sh "Open https://example.com and report the page title"
Grounding
对仓库随附的三个样本运行直接 grounding 评估:
MODEL_URL=http://127.0.0.1:8000/v1 \
MODEL_NAME=UI-Venus-2 \
bash scripts/grounding.sh
CAPTCHA
对随附的 CAPTCHA 图像运行推理,并将解析出的 JSON 和可视化结果保存到 results/captcha/:
MODEL_URL=http://127.0.0.1:8000/v1 \
MODEL_NAME=UI-Venus-2 \
bash scripts/captcha.sh
📊 基准测试结果
移动端
| 模型 | MobileGym | VenusBench‑Mobile | AndroidWorld | MobileWorld | KnowUBench | MemGUI |
|---|---|---|---|---|---|---|
| 通用 VLM | ||||||
| Qwen3.5‑9B | 9.0* | 15.3* | 57.8 | 18.0 (18.0)* | 33.3 | 6.2* |
| Qwen3.6‑27B | 24.6* | 28.0* | 70.3 | 36.8 (41.9)* | - | 25.7* |
| Claude‑Opus‑4.6 | - | 36.5* | - | 44.5 | - | - |
| Kimi‑K2.6 | 38.7* | 31.2* | - | 55.6 | - | 39.1 |
| Kimi‑K3 | - | - | - | 74.4 | - | - |
| Seed‑2.0‑Pro | 52.0 | 20.1* | - | 63.2 | 51.6 | 65.6* |
| Seed‑2.1‑Pro | - | - | - | 73.2 | - | - |
| GPT‑5.6‑Sol | - | - | - | 70.1 | - | - |
| GUI 专用模型 | ||||||
| UI‑Venus‑1.5‑8B | 18.4* | 16.1 | 73.7 | 22.2* | 26.0 | 3.9* |
| UI‑Venus‑1.5‑30B‑A3B | 21.5* | 21.5 | 77.6 | 17.1 | - | 10.9* |
| GUI‑Owl‑1.5‑32B‑Instruct | 20.3* | - | 69.8 | 43.9 | - | 10.9 |
| MAI‑UI‑8B | 21.5* | 12.7 | 70.7 | 27.5 | 26.0 | 17.2* |
| Qwen‑UI‑Agent‑27B | - | - | - | 82.1 (85.5) | - | - |
| Ours | ||||||
| UI‑Venus‑2‑9B | 52.7 | 46.5 | 80.2 | 65.8 (75.2) | 56.5 | 62.6 |
| UI‑Venus‑2‑27B | 60.5 | 48.7 | 84.0 | 76.1 (82.9) | 59.7 | 70.3 |
在多个移动端 GUI 基准上的性能对比。VenusBench-Mobile 报告的是其 149 项任务主任务池上的成功率。对于 MobileWorld,我们报告的是 50 步设置下 117 项任务的仅 GUI 成功率;括号中的数值(如有)使用 100 步。MemGUI 报告的是 Main Results Pass@1。
*表示由我们评估或复现的基线结果。
Computer
|
| |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
在计算机使用智能体基准测试上的性能对比:OSWorld-Verified 和 DeskCraft(左图),以及 OSWorld 2.0(右图),均在官方 150 步预算、108 个任务的设定下进行。所报告的 OSWorld-Verified 基线在其引用来源中使用的是 361 个任务的设定,且可能采用针对特定模型的动作脚手架。对于 DeskCraft,我们报告的是作者在 Standard 和 Interactive 两个划分的 538 个任务并集上评估的汇总结果,这与该基准官方的按划分报告方式不同。OSWorld 2.0 结果报告的是官方的 Binary Accuracy 和 Partial Score 指标;基线取自官方排行榜,可能采用了针对特定模型的工具设置,对于有多个官方条目的模型,其推理强度设置在括号中标注。
*表示由我们评估的基线结果。
浏览器
| 模型 | WebVoyager | Online‑Mind2Web | REAL | Odysseys | |
|---|---|---|---|---|---|
| 平均 | Perfect | ||||
| 通用 VLM | |||||
| Qwen3.5‑9B | 46.9* | 27.3* | 18.2* | 42.6* | 13.5* |
| Qwen3.5‑4B | - | - | - | 42.9 | 10.7 |
| Qwen3.6‑27B | 84.3* | 55.3* | 27.3* | 39.5* | 18.5* |
| OpenAI Operator | 87.0 | 61.3 | - | - | - |
| GPT‑5 (SoM) | 90.6 | - | - | - | - |
| GPT‑5.4 | - | - | - | 55.4 | 33.5 |
| Seed2.0 Pro | 85.1* | 68.5* | 74.4* | 60.2* | 30.1* |
| GLM‑5V‑Turbo | 88.5 | - | - | - | - |
| Claude Opus 4.6 | 88.0 | - | - | 68.9 | 44.5 |
| Claude‑Sonnet‑4.6 | - | - | - | 49.8 | 31.0 |
| Kimi‑K2.6 | 76.8* | - | 74.4* | - | - |
| GUI 专用模型 | |||||
| UI‑TARS‑1.5 | 84.8 | 75.8 | - | - | - |
| UI‑Venus‑1.5‑30B‑A3B | 76.0 | - | 38.0* | - | - |
| GUI‑Owl‑1.5‑32B‑Thinking | 82.1 | - | 44.6* | - | - |
| MolmoWeb‑8B | 78.2 | 35.3 | - | - | - |
| Fara1.5‑4B | 80.8 | - | - | - | - |
| Fara1.5‑9B | 86.6 | 63.4 | - | - | - |
| Fara1.5‑27B | 89.3 | 72.3 | - | - | - |
| Ours | |||||
| UI‑Venus‑2‑9B | 90.8 | 74.0 | 76.9 | 77.3 | 62.0 |
| UI‑Venus‑2‑27B | 93.4 | 78.3 | 80.2 | 80.4 | 66.3 |
在四个实时网页基准上的性能对比:WebVoyager、Online-Mind2Web、REAL 和 Odysseys。Fara1.5 和 GPT-5 (SoM) 的 WebVoyager 条目采用更新后的 595 任务、100 步稳健协议,并取三次运行的平均值;实时站点状态可能因评估日期而异。对于 Odysseys,我们同时报告平均评分(Avg.)和满分评分(Perfect)。
*表示我们复现的结果。
Grounding
| 模型 | VenusBench‑GD | ScreenSpot‑Pro | OSWorld‑G‑R | UI‑Vision |
|---|---|---|---|---|
| 通用 VLM | ||||
| Qwen 3.7 Plus | 75.2* | 68.9 | 78.2 | 68.0 |
| Seed 2.1 Pro | 73.9* | 65.3 | 78.0 | 62.0 |
| Kimi‑K2.6 | 73.1* | 52.0* | 69.7* | 51.7* |
| Qwen3.6‑27B | 67.7* | 65.2* | 76.9* | 58.3* |
| GUI 专用模型 | ||||
| UI‑Venus‑Ground‑72B | 70.2 | 61.9 | 69.5 | 36.8 |
| Holo2‑30B‑A3B | 59.5* | 66.1 | 76.1 | 40.9* |
| Step‑GUI‑4B | 54.6* | 60.0 | 66.9 | 30.0* |
| MAI‑UI‑8B | 65.2* | 65.8 | 68.6 | 40.7 |
| MAI‑UI‑32B | - | 67.9 | 73.9 | 47.1 |
| UI‑Venus‑1.5‑30B‑A3B | 75.0 | 69.6 | 76.4 | 54.7 |
| Qwen‑UI‑Agent‑27B | - | 76.6 | 78.5 | 70.0 |
| Ours | ||||
| UI‑Venus‑2‑9B | 77.1 | 73.0 | 78.5 | 53.2 |
| UI‑Venus‑2‑27B | 80.1 | 74.1 | 79.1 | 66.9 |
在多个 Grounding 基准上的性能对比。VenusBench-GD 报告的是英文指令下的微平均 point-in-box 准确率。
*表示由我们评估或复现的基线。
CAPTCHA
| 模型 | VenusBench‑CAPTCHA | MCA‑Bench | Spatial‑CAPTCHA‑Bench | NextGen‑CAPTCHAs | Open CaptchaWorld |
|---|---|---|---|---|---|
| 通用 VLM | |||||
| Qwen3.5‑9B | 28.3 | 30.4 | 4.9 | 2.8 | 36.4 |
| Qwen3.6‑27B | 53.0 | 51.7 | 31.0 | 14.1 | 47.7 |
| Seed‑2.0‑Pro | 47.9 | 36.5 | 43.8 | 20.4 | 55.6 |
| Kimi‑K2.6 | 39.7 | 38.7 | 24.8 | 7.2 | 47.8 |
| Claude‑Opus‑4.6 | 16.0 | 25.9 | 9.5 | 2.8 | 23.3 |
| Ours | |||||
| UI‑Venus‑2‑9B | 78.1 | 75.7 | 42.8 | 47.6 | 50.7 |
| UI‑Venus‑2‑27B | 79.9 | 79.6 | 48.6 | 54.5 | 56.3 |
各 CAPTCHA 基准上的性能对比。所有结果均为 Pass@1 百分比,越高越好。我们在 VenusBench-CAPTCHA 以及四个公开基准上进行评估:MCA-Bench、Spatial-CAPTCHA-Bench、NextGen-CAPTCHAs 和 Open CaptchaWorld。我们使用了 1,000 个抽样的 MCA-Bench 样本、15 种 NextGen-CAPTCHAs 任务类型以及 16 种 Open CaptchaWorld 任务类型。
📬 联系方式
如有任何问题或合作意向,请联系维护者。
📚 引用
# UI-Venus-2
@misc{venusteam2026uivenus2technicalreport,
title={UI-Venus-2 Technical Report},
author={Venus Team and Zhuohan Cai and Haoxing Chen and Jiaxuan Chen and Weizhi Chen and Changlong Gao and Zhangxuan Gu and Yuan Guo and Yusong Hu and Jianrong Jiang and Jianguo Li and Runze Li and Jinzhen Lin and Zhenyu Ma and Changhua Meng and Han Peng and Xinyu Qiu and Shuheng Shen and Zhongyi Shui and Weiqiang Wang and Ming Wen and Zhuoer Xu and Hang Yan and Kaiwen Yang and Ruilin Yao and Nanjun Yu and Zhengwen Zeng and Lianrui Zhang and Yunzhu Zhang and Zhe Zhao and Beitong Zhou},
year={2026},
eprint={2609.00028},
archivePrefix={arXiv},
primaryClass={cs.AI},
url={https://arxiv.org/abs/2609.00028},
}
# UI-Venus-1.5
@misc{venusteam2026uivenus15technicalreport,
title={UI-Venus-1.5 Technical Report},
author={Venus-Team and Changlong Gao and Zhangxuan Gu and Yulin Liu and Xinyu Qiu and Shuheng Shen and Yue Wen and Tianyu Xia and Zhenyu Xu and Zhengwen Zeng and Beitong Zhou and Xingran Zhou and Weizhi Chen and Sunhao Dai and Jingya Dou and Yichen Gong and Yuan Guo and Zhenlin Guo and Feng Li and Qian Li and Jinzhen Lin and Yuqi Zhou and Linchao Zhu and Liang Chen and Zhenyu Guo and Changhua Meng and Weiqiang Wang},
year={2026},
eprint={2602.09082},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2602.09082},
}
# UI-Venus
@misc{gu2025uivenustechnicalreportbuilding,
title={UI-Venus Technical Report: Building High-performance UI Agents with RFT},
author={Zhangxuan Gu and Zhengwen Zeng and Zhenyu Xu and others},
year={2025},
eprint={2508.10833},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2508.10833},
}
⚖️ 许可证
本项目仅供研究与教育用途。
A general-purpose foundation GUI agent for mobile apps, web platforms, and desktop operating systems — scaling environments, tasks, and feedback jointly within one closed-loop perception–reasoning–action agent.
🌟 What's New in UI-Venus-2
Compared with UI-Venus-1.5, we introduce:
- 📱 Scaled multilingual mobile environments: A substantially expanded executable mobile pool covering Chinese and English app ecosystems, paired with a deep-research-driven query-generation strategy that grounds task queries in real application functionality — improving the accuracy, validity, and executability of generated instructions.
- 🖥️ Computer use, built from the ground up: Dedicated desktop operating-system capabilities constructed from scratch through computer-use data collection and task-specific training, extending the UI-Venus family to mobile, web, and OS interaction in one unified end-to-end agent.
- 🎯 Keypoint-grounded verification: Task completion is judged on task-relevant visual keypoints rather than a coarse holistic look at the final screen, with multi-model voting aggregating heterogeneous judges — reducing single-judge bias and making the reward signal robust to reward hacking.
- 🔄 Verification-augmented reflection: Verified feedback is distilled back into training as reflection supervision, so the agent can distinguish partial progress from true completion, avoid premature termination caused by observation misinterpretation, and recover during long-horizon interaction.
📈 UI-Venus-2 Benchmark Performance
Performances of UI-Venus-2 on GUI-agent benchmarks. Each panel compares UI-Venus-2-27B and UI-Venus-2-9B with some selected strong baselines. We favor standalone end-to-end systems evaluated on the closest available task subset and step budget; source-reported action scaffolds may still differ. MobileWorld uses GUI-only success rate on 117 tasks with 50 steps, WebVoyager uses the refreshed 595-task split, Odysseys uses average rubric score over 200 tasks, VenusBench-CAPTCHA uses micro Pass@1 over all 219 examples, and VenusBench-GD uses English-instruction micro-average accuracy. “*” denotes results reproduced by us.
📰 News
- [2026/09] We release VenusBench-CAPTCHA, a comprehensive benchmark for evaluating GUI agents on diverse CAPTCHA tasks. See branch VenusBench-CAPTCHA.
- [2026/08] We release UI-Venus-2, a 9B/27B general-purpose foundation GUI agent that unifies mobile, web, and desktop interaction with scaled multilingual environments, keypoint-grounded verification, and verification-augmented reflection.
- [2026/02] We release UI-Venus-1.5, an end-to-end GUI Agent designed for robust real-world applications.
- [2026/02] We release VenusBench-Mobile, a challenging online benchmark for mobile GUI agents. See branch VenusBench-Mobile.
- [2025/12] We release VenusBench-GD, a comprehensive multi-platform GUI grounding benchmark. See branch VenusBench-GD.
- [2025/8] We release UI-Venus, the first version of our UI agent model.
🧭 Overview
✨ Demo
🛠 Venus Framework
We provide two frameworks for running agents in real environments:
| Framework | Description | Documentation |
|---|---|---|
| Mobile Framework | Android/ADB agent framework for single-task execution, multi-device batch execution, trajectory recording, and reflection. | English · 中文 |
| Browser Plugin | Chrome Side Panel extension that connects UI-Venus to the active browser tab and executes browser tasks interactively. | English · 中文 |
See the Venus Framework overview for the directory layout and entry points. The lightweight domain examples below can be used without either framework unless their individual requirements state otherwise.
🚀 Quick Start
Installation
conda create -n ui-venus-2 python=3.11 -y
conda activate ui-venus-2
pip install -r requirements.txt
Python 3.10 or newer is required. All commands below are executed from the repository root. Configure the OpenAI-compatible model service through MODEL_URL, MODEL_NAME, and either API_KEY or MODEL_API_KEY; the same values can also be edited at the top of each domain script. API_KEY takes precedence when both key variables are set. Set MODEL_NAME to the served name of your UI-Venus-2 9B or 27B model.
Mobile
Run multi-turn inference over the included prerecorded screenshot sequence. N_IMG controls how many recent historical screenshots are retained:
MODEL_URL=http://127.0.0.1:8000/v1 \
MODEL_NAME=UI-Venus-2 \
N_IMG=2 \
bash scripts/mobile.sh
This example performs model inference only and does not execute actions on a device. Use the Mobile Framework above for real-device ADB automation.
Mobile multi-turn example and input/output format
Computer
Run multi-turn Computer inference over a prerecorded desktop screenshot sequence:
MODEL_URL=http://127.0.0.1:8000/v1 \
MODEL_NAME=UI-Venus-2 \
N_IMG=2 \
bash scripts/computer.sh
The default command uses the included desktop screenshot sample. The standalone example validates and normalizes model actions but does not execute them on the host. It has no runtime dependency on OSWorld.
Computer multi-turn example and action format
Browser
Start Chrome with a CDP port as described in the domain documentation, then run one natural-language browser task:
MODEL_URL=http://127.0.0.1:8000/v1 \
MODEL_NAME=UI-Venus-2 \
bash scripts/browser.sh "Open https://example.com and report the page title"
Grounding
Run the direct grounding evaluation on the three samples included with the repository:
MODEL_URL=http://127.0.0.1:8000/v1 \
MODEL_NAME=UI-Venus-2 \
bash scripts/grounding.sh
Grounding evaluation, smoke test, and benchmark configuration
CAPTCHA
Run inference on the included CAPTCHA image and save the parsed JSON and visualization under results/captcha/:
MODEL_URL=http://127.0.0.1:8000/v1 \
MODEL_NAME=UI-Venus-2 \
bash scripts/captcha.sh
CAPTCHA usage, prompt, action format, and visualization
📊 Benchmark Results
Mobile
| Models | MobileGym | VenusBench‑Mobile | AndroidWorld | MobileWorld | KnowUBench | MemGUI |
|---|---|---|---|---|---|---|
| General VLMs | ||||||
| Qwen3.5‑9B | 9.0* | 15.3* | 57.8 | 18.0 (18.0)* | 33.3 | 6.2* |
| Qwen3.6‑27B | 24.6* | 28.0* | 70.3 | 36.8 (41.9)* | - | 25.7* |
| Claude‑Opus‑4.6 | - | 36.5* | - | 44.5 | - | - |
| Kimi‑K2.6 | 38.7* | 31.2* | - | 55.6 | - | 39.1 |
| Kimi‑K3 | - | - | - | 74.4 | - | - |
| Seed‑2.0‑Pro | 52.0 | 20.1* | - | 63.2 | 51.6 | 65.6* |
| Seed‑2.1‑Pro | - | - | - | 73.2 | - | - |
| GPT‑5.6‑Sol | - | - | - | 70.1 | - | - |
| GUI‑specific Models | ||||||
| UI‑Venus‑1.5‑8B | 18.4* | 16.1 | 73.7 | 22.2* | 26.0 | 3.9* |
| UI‑Venus‑1.5‑30B‑A3B | 21.5* | 21.5 | 77.6 | 17.1 | - | 10.9* |
| GUI‑Owl‑1.5‑32B‑Instruct | 20.3* | - | 69.8 | 43.9 | - | 10.9 |
| MAI‑UI‑8B | 21.5* | 12.7 | 70.7 | 27.5 | 26.0 | 17.2* |
| Qwen‑UI‑Agent‑27B | - | - | - | 82.1 (85.5) | - | - |
| Ours | ||||||
| UI‑Venus‑2‑9B | 52.7 | 46.5 | 80.2 | 65.8 (75.2) | 56.5 | 62.6 |
| UI‑Venus‑2‑27B | 60.5 | 48.7 | 84.0 | 76.1 (82.9) | 59.7 | 70.3 |
Performance comparison on various mobile GUI benchmarks. VenusBench-Mobile reports success rate on its 149-task primary pool. For MobileWorld, we report GUI-only success rate on 117 tasks under the 50-step setting; values in parentheses, when available, use 100 steps. MemGUI reports Main Results Pass@1.
*denotes baseline results evaluated or reproduced by us.
Computer
|
| |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Performance comparison on computer-use agent benchmarks: OSWorld-Verified and DeskCraft (left), and OSWorld 2.0 (right) under the official 150-step budget with 108 tasks. Reported OSWorld-Verified baselines use the 361-task setting in their cited source and may use model-specific action scaffolds. For DeskCraft, we report an author-evaluated aggregate over the 538-task union of the Standard and Interactive splits, which differs from the benchmark's official split-level reporting. OSWorld 2.0 results report the official Binary Accuracy and Partial Score metrics; baselines are taken from the official leaderboard, possibly with model-specific tool settings, and the reasoning-effort setting is labeled in parentheses for models with multiple official entries.
*indicates baseline results evaluated by us.
Browser
| Models | WebVoyager | Online‑Mind2Web | REAL | Odysseys | |
|---|---|---|---|---|---|
| Avg. | Perfect | ||||
| General VLMs | |||||
| Qwen3.5‑9B | 46.9* | 27.3* | 18.2* | 42.6* | 13.5* |
| Qwen3.5‑4B | - | - | - | 42.9 | 10.7 |
| Qwen3.6‑27B | 84.3* | 55.3* | 27.3* | 39.5* | 18.5* |
| OpenAI Operator | 87.0 | 61.3 | - | - | - |
| GPT‑5 (SoM) | 90.6 | - | - | - | - |
| GPT‑5.4 | - | - | - | 55.4 | 33.5 |
| Seed2.0 Pro | 85.1* | 68.5* | 74.4* | 60.2* | 30.1* |
| GLM‑5V‑Turbo | 88.5 | - | - | - | - |
| Claude Opus 4.6 | 88.0 | - | - | 68.9 | 44.5 |
| Claude‑Sonnet‑4.6 | - | - | - | 49.8 | 31.0 |
| Kimi‑K2.6 | 76.8* | - | 74.4* | - | - |
| GUI‑specific Models | |||||
| UI‑TARS‑1.5 | 84.8 | 75.8 | - | - | - |
| UI‑Venus‑1.5‑30B‑A3B | 76.0 | - | 38.0* | - | - |
| GUI‑Owl‑1.5‑32B‑Thinking | 82.1 | - | 44.6* | - | - |
| MolmoWeb‑8B | 78.2 | 35.3 | - | - | - |
| Fara1.5‑4B | 80.8 | - | - | - | - |
| Fara1.5‑9B | 86.6 | 63.4 | - | - | - |
| Fara1.5‑27B | 89.3 | 72.3 | - | - | - |
| Ours | |||||
| UI‑Venus‑2‑9B | 90.8 | 74.0 | 76.9 | 77.3 | 62.0 |
| UI‑Venus‑2‑27B | 93.4 | 78.3 | 80.2 | 80.4 | 66.3 |
Performance comparison on four live-web benchmarks: WebVoyager, Online-Mind2Web, REAL, and Odysseys. The Fara1.5 and GPT-5 (SoM) WebVoyager entries use the refreshed 595-task, 100-step robust protocol and are averaged over three runs; live-site states may vary by evaluation date. For Odysseys, we report both the averaged rubric score (Avg.) and the perfect rubric score (Perfect).
*indicates our reproduced results.
Grounding
| Models | VenusBench‑GD | ScreenSpot‑Pro | OSWorld‑G‑R | UI‑Vision |
|---|---|---|---|---|
| General VLMs | ||||
| Qwen 3.7 Plus | 75.2* | 68.9 | 78.2 | 68.0 |
| Seed 2.1 Pro | 73.9* | 65.3 | 78.0 | 62.0 |
| Kimi‑K2.6 | 73.1* | 52.0* | 69.7* | 51.7* |
| Qwen3.6‑27B | 67.7* | 65.2* | 76.9* | 58.3* |
| GUI‑specific Models | ||||
| UI‑Venus‑Ground‑72B | 70.2 | 61.9 | 69.5 | 36.8 |
| Holo2‑30B‑A3B | 59.5* | 66.1 | 76.1 | 40.9* |
| Step‑GUI‑4B | 54.6* | 60.0 | 66.9 | 30.0* |
| MAI‑UI‑8B | 65.2* | 65.8 | 68.6 | 40.7 |
| MAI‑UI‑32B | - | 67.9 | 73.9 | 47.1 |
| UI‑Venus‑1.5‑30B‑A3B | 75.0 | 69.6 | 76.4 | 54.7 |
| Qwen‑UI‑Agent‑27B | - | 76.6 | 78.5 | 70.0 |
| Ours | ||||
| UI‑Venus‑2‑9B | 77.1 | 73.0 | 78.5 | 53.2 |
| UI‑Venus‑2‑27B | 80.1 | 74.1 | 79.1 | 66.9 |
Performance comparison on various Grounding Benchmarks. VenusBench-GD reports English-instruction micro-average point-in-box accuracy.
*indicates baselines evaluated or reproduced by us.
CAPTCHA
| Models | VenusBench‑CAPTCHA | MCA‑Bench | Spatial‑CAPTCHA‑Bench | NextGen‑CAPTCHAs | Open CaptchaWorld |
|---|---|---|---|---|---|
| General VLMs | |||||
| Qwen3.5‑9B | 28.3 | 30.4 | 4.9 | 2.8 | 36.4 |
| Qwen3.6‑27B | 53.0 | 51.7 | 31.0 | 14.1 | 47.7 |
| Seed‑2.0‑Pro | 47.9 | 36.5 | 43.8 | 20.4 | 55.6 |
| Kimi‑K2.6 | 39.7 | 38.7 | 24.8 | 7.2 | 47.8 |
| Claude‑Opus‑4.6 | 16.0 | 25.9 | 9.5 | 2.8 | 23.3 |
| Ours | |||||
| UI‑Venus‑2‑9B | 78.1 | 75.7 | 42.8 | 47.6 | 50.7 |
| UI‑Venus‑2‑27B | 79.9 | 79.6 | 48.6 | 54.5 | 56.3 |
Performance Comparison across CAPTCHA Benchmarks. All results are Pass@1 percentages, and higher is better. We evaluate on VenusBench-CAPTCHA and four public benchmarks: MCA-Bench, Spatial-CAPTCHA-Bench, NextGen-CAPTCHAs, and Open CaptchaWorld. We use 1,000 sampled MCA-Bench examples, 15 NextGen-CAPTCHAs task types, and 16 Open CaptchaWorld task types.
📬 Contact
For any questions or collaboration, please contact the maintainers.
📚 Citation
# UI-Venus-2
@misc{venusteam2026uivenus2technicalreport,
title={UI-Venus-2 Technical Report},
author={Venus Team and Zhuohan Cai and Haoxing Chen and Jiaxuan Chen and Weizhi Chen and Changlong Gao and Zhangxuan Gu and Yuan Guo and Yusong Hu and Jianrong Jiang and Jianguo Li and Runze Li and Jinzhen Lin and Zhenyu Ma and Changhua Meng and Han Peng and Xinyu Qiu and Shuheng Shen and Zhongyi Shui and Weiqiang Wang and Ming Wen and Zhuoer Xu and Hang Yan and Kaiwen Yang and Ruilin Yao and Nanjun Yu and Zhengwen Zeng and Lianrui Zhang and Yunzhu Zhang and Zhe Zhao and Beitong Zhou},
year={2026},
eprint={2609.00028},
archivePrefix={arXiv},
primaryClass={cs.AI},
url={https://arxiv.org/abs/2609.00028},
}
# UI-Venus-1.5
@misc{venusteam2026uivenus15technicalreport,
title={UI-Venus-1.5 Technical Report},
author={Venus-Team and Changlong Gao and Zhangxuan Gu and Yulin Liu and Xinyu Qiu and Shuheng Shen and Yue Wen and Tianyu Xia and Zhenyu Xu and Zhengwen Zeng and Beitong Zhou and Xingran Zhou and Weizhi Chen and Sunhao Dai and Jingya Dou and Yichen Gong and Yuan Guo and Zhenlin Guo and Feng Li and Qian Li and Jinzhen Lin and Yuqi Zhou and Linchao Zhu and Liang Chen and Zhenyu Guo and Changhua Meng and Weiqiang Wang},
year={2026},
eprint={2602.09082},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2602.09082},
}
# UI-Venus
@misc{gu2025uivenustechnicalreportbuilding,
title={UI-Venus Technical Report: Building High-performance UI Agents with RFT},
author={Zhangxuan Gu and Zhengwen Zeng and Zhenyu Xu and others},
year={2025},
eprint={2508.10833},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2508.10833},
}
⚖️ License
This project is for research and educational purposes only.