UI-Venus-2 scales GUI agent to 170+ apps and desktop OSes
Researchers behind UI-Venus-2 describe it as a general-purpose foundation GUI agent meant to operate across mobile, web and desktop environments through a single closed-loop reasoning-action framework. The paper argues that most GUI agents look strong on benchmarks but struggle once deployed in the real world, and traces that gap to three causes: limited environment coverage, task construction that breaks easily, and reward verification that cannot be trusted.
To close that gap, the team says they scaled three parts of the system together. Environment coverage was expanded to more than 170 multilingual mobile apps plus native desktop operating systems, broader than the testbeds typical of prior GUI agents. Task generation runs through what the authors call a deep-research pipeline for function-grounded instruction generation, meant to produce tasks tied to what an app's functions actually do rather than generic scripted prompts. Verification adds trace-level and sample-level evaluators that use visual keypoints and multi-model voting, intended to give the reinforcement-learning training loop signals it can rely on.
The paper also describes safety-aware mechanisms built in to keep consequential actions, the kind that could delete data, send messages or make purchases, under controlled execution rather than letting the agent act on them freely. The authors present UI-Venus-2 as a capable, efficient and open-source foundation intended to push the field toward agents that are more generalizable, verifiable and self-reflective in real-world use.
The available text is the paper's abstract. It does not name the authors or their institutions, does not give benchmark scores or comparisons with earlier UI-Venus versions or rival GUI agents, does not specify which desktop operating systems are covered beyond 'native desktop operating systems,' and does not give model size or training data volume.
Key facts
- Environment coverage expanded to more than 170 multilingual mobile apps plus native desktop operating systems.
- Task generation uses a deep-research pipeline for function-grounded instruction generation rather than generic scripted prompts.
- Verification combines trace-level and sample-level evaluators with visual keypoints and multi-model voting to produce reliable reinforcement-learning signals.
- Safety-aware mechanisms are built in to keep consequential actions under controlled execution.
- The authors describe UI-Venus-2 as a capable, efficient and open-source foundation model.
Why it matters
Most GUI agents look strong in benchmark demos, then break once asked to operate real apps outside a curated test set. UI-Venus-2's authors position it as an attempt to close that gap directly, by scaling three things that usually stay narrow: how many real environments the agent is tested against, how its training tasks get written, and how its actions get verified during training. That combination, rather than a single model trick, is the paper's central claim to being closer to a deployable agent than a benchmark exercise.
Who it affects
Anyone building on GUI automation, robotic process automation, mobile QA testing, or agentic assistants that need to click through real apps rather than call APIs stands to gain from a foundation model with broader environment coverage. Because the authors describe UI-Venus-2 as open-source, developers integrating GUI agents into products could adopt it directly instead of building environment coverage themselves.
How to use it
The paper offers UI-Venus-2 as an open-source foundation for GUI agent work covering mobile, web and desktop targets. The abstract does not give a release date, license terms, model size, or where the weights or code are hosted, so anyone wanting to use it will need to check the paper or its accompanying repository for those specifics.
How solid is it
This is a technical report on arXiv, so the description above comes from the authors themselves and is not independently benchmarked here. The abstract does not include benchmark scores, accuracy numbers, or head-to-head comparisons with prior GUI agents, including earlier UI-Venus versions, so the practical scale of the improvement is not demonstrable from what is available.
Risks and caveats
The source text is the paper's abstract and does not name the authors, their institutions, the release timeline, the model size, the training data volume, or which specific desktop operating systems are covered beyond 'native desktop operating systems.' Claims about safety-aware mechanisms and reliable reinforcement-learning signals are the authors' own framing of their work, not independently verified results.
“Transitioning from benchmark-oriented models to dependable real-world applications remains challenging due to limited environment coverage, brittle task construction, and unreliable reward verification.”
— UI-Venus-2 paper abstract