inclusionAI/UI-Venus-2-9B

UI-Venus-2-9B

UI-Venus-2 is a general-purpose foundation GUI agent designed to operate across mobile applications, web platforms, and desktop operating systems through a unified closed-loop reasoning–action framework: the agent observes the current interface, reasons about the task state, executes an action, and incorporates environmental feedback into its next decision.

To bridge the gap from benchmark-oriented models toward dependable real-world applications, UI-Venus-2 jointly scales three critical dimensions:

• Environments — expanding coverage to 170+ multilingual mobile apps (100+ Chinese and 70+ English), 50k+ live websites, and native desktop OS with 50+ professional applications;
• Tasks — a deep-research pipeline that grounds generated instructions in the actual functionality of each application, improving the accuracy, validity, and executability of task queries;
• Verification — trace-level and sample-level evaluators based on task-relevant visual keypoints and multi-model voting, providing reliable RL reward signals that are robust to reward hacking.

UI-Venus-2 further integrates safety-aware mechanisms to ensure controlled execution of consequential actions, bringing the attack success rate on OSBlind down from 90%+ (typical of prior GUI agents) to 12.3%.

This repository releases the full-parameter weights of UI-Venus-2-9B, initialized from Qwen3.5-9B, which achieves near state-of-the-art performance among models of comparable scale across GUI grounding, mobile, web, computer-use, and CAPTCHA benchmarks.

Highlights

• Scaled multilingual mobile-use environments with reliable task generation. A substantially expanded executable mobile pool covering Chinese and English app ecosystems, paired with a deep-research-driven query-generation strategy grounded in real application functionality.
• Computer-use capability 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, and recover during long-horizon interaction.

Training Pipeline

UI-Venus-2 is trained on a deliberately complementary mixture of five task families — Grounding, CAPTCHA, Mobile, Web, and Computer — through three stages:

1. Multimodal Mid-Training — large-scale trajectory-based mid-training over simulated mobile, web, and OS environments, with human–discriminator collaborative verification filtering invalid or ambiguous interactions, plus programmatically synthesized grounding and CAPTCHA supervision. 2. Offline RL — step-level RL trajectories for Mobile, OS, and Web optimize state-aware action selection and execution reliability, while verified CAPTCHA and grounding instances embedded in realistic interfaces teach precise localization under visual clutter. 3. Multi-teacher On-policy Distillation — domain-specialized teachers are consolidated into a single unified policy that preserves the broad multimodal reasoning of the base model while composing spatial grounding, verified interaction, and long-horizon navigation.

The training data is produced by a deep-research-driven task-generation pipeline and collected at scale across mobile (ADB, 200+ apps), browser (Playwright, 50k+ websites), and computer (PyAutoGUI / VM, 50+ desktop applications) environments:

Quick Start

Serving with vLLM

pip install -U vllm

vllm serve inclusionAI/UI-Venus-2-9B \
--port 8000 \
--tensor-parallel-size 1 \
--max-model-len 262144 \
--reasoning-parser qwen3

The server exposes an OpenAI-compatible API at http://localhost:8000/v1.

For the full agent scaffolding — grounding / navigation prompts, the action space, ADB / Playwright / VM deployment, and evaluation infrastructure — please refer to our GitHub repository.

Results

Key results per domain, with strong published baselines for context. A dash denotes a result not currently reported. Best result per benchmark in bold. Full comparison tables and metric definitions are available in the technical report.

Mobile Use

Success rate (%) on online, interactive mobile-agent benchmarks.

| Model | AndroidWorld | MobileWorld | VenusBench-Mobile | |---|:---:|:---:|:---:| | UI-Venus-2-9B (ours) | 80.2 | 65.8 | 46.5 | | UI-Venus-2-27B | 84.0 | 76.1 | 48.7 | | UI-Venus-1.5-30B-A3B | 77.6 | — | 21.5 | | Qwen-UI-Agent-27B | — | 82.1 | — | | Qwen3.6-27B | 70.3 | — | 28.0 | | MAI-UI-8B | 70.7 | — | — | | GUI-Owl-1.5-32B-Instruct | 69.8 | — | — | | Claude Opus 4.6 | — | 44.5 | 36.5 | | Kimi K2.6 | — | 55.6 | 31.2 | | Seed2.0 Pro | — | 63.2 | — |

Computer Use

Success rate (%) on open-ended tasks in real desktop environments.

| Model | OSWorld-Verified | DeskCraft | |---|:---:|:---:| | UI-Venus-2-9B (ours) | 70.8 | 48.0 | | UI-Venus-2-27B | 80.5 | 55.5 | | Qwen-UI-Agent-27B | 79.5 | — | | Kimi K2.6 | 73.1 | 41.4 | | Qwen3.6-27B | 62.0 | 28.7 | | GUI-Owl-1.5-32B-Instruct | 56.5 | — | | Seed2.0 Pro | — | 40.0 | | Qwen3.5-9B | — | 14.6 |

Web Navigation

Success rate (%) on end-to-end web navigation; Odysseys reports the average rubric score (%).

| Model | WebVoyager | Online-Mind2Web | REAL | Odysseys | |---|:---:|:---:|:---:|:---:| | UI-Venus-2-9B (ours) | 90.8 | 74.0 | 76.9 | 77.3 | | UI-Venus-2-27B | 93.4 | 78.3 | 80.2 | 80.4 | | UI-Venus-1.5-30B-A3B | — | — | 38.0 | — | | Fara-1.5-27B | 89.3 | 72.3 | — | — | | GLM-5V-Turbo | 88.5 | — | — | — | | Claude Opus 4.6 | 88.0 | — | — | 68.9 | | Seed-2.0 | 85.1 | 68.5 | 74.4 | 60.2 | | UI-TARS-1.5 | — | 75.8 | — | — | | OpenAI Operator | — | 61.3 | — | — | | Kimi K2.6 | — | — | 74.4 | — |

GUI Grounding

Accuracy (%) on element localization across high-resolution professional software,...