UI-Venus-2

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 the executable mobile pool to 170+ multilingual apps (100+ Chinese and 70+ English), alongside a curated web pool of 4,000+ domains across 19 categories and native desktop operating systems;
• 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 is evaluated for consequential-action safety on OSHarm and OSBlind. UI-Venus-2-9B achieves an Attack Success Rate (ASR, lower is better) of 11.3% on OSHarm and 48.8% on OSBlind, compared with 25.3% and 79.4%, respectively, for its Qwen3.5-9B base model.

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.
• Multi-teacher on-policy distillation. Domain-specialized teachers are consolidated with structured, action-aware supervision that concentrates the distillation signal on the executable behaviors governing environment 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 training over simulated mobile, web, and OS environments, with navigation data as the dominant component and human–discriminator collaborative verification filtering invalid or ambiguous interactions.
2. Offline RL — independent step-level optimization across Grounding, CAPTCHA, Mobile, Web, and Computer tasks develops domain-specialized policies.
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 spanning mobile, browser, and computer environments. The active report describes a curated web pool of 4,000+ domains across 19 categories, seeded with 45,000 tasks selected from InSTA-150k-v3; synthetic grounding and CAPTCHA records are exported for task-specific SFT and RL.

Quick Start

Serving with vLLM

The following is the intended serving configuration. It was not live-canary-validated as part of this model-card update, so pin and verify a vLLM version appropriate for Qwen3.5 before deployment.

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.

Inference Configuration

The evaluation settings depend on the task family:

• General agentic tasks: use temperature 1.0, enable reasoning, and retain the complete reasoning history in the multi-step context.
• GUI grounding: disable reasoning and use temperature 0 for direct point localization.
• CAPTCHA: use the CAPTCHA-specific prompt and a parser/executor that supports multiple actions in one model turn.

Starting the vLLM server alone does not provide the complete closed-loop GUI agent. Reference prompts, action parsers, and runnable examples for Mobile, Computer, Browser, Grounding, and CAPTCHA are available in the UI-Venus code repository. Environment coverage differs by domain: the included Mobile and Computer examples support multi-turn inference over prerecorded screenshots, while real-device ADB automation, browser interaction, and benchmark-specific setup are documented separately in the repository.

Release Status

• Model weights: included in this repository.
• Code: available in the UI-Venus GitHub repository.
• Project page: available at ui-venus.github.io/UI-Venus-2.
• **Technical report a