UI-Venus-2: The GUI Agent That Actually Works in the Wild
UI-Venus-2 is the first general-purpose GUI agent that can reliably automate real-world digital tasks across 170+ apps and operating systems.
Why it earned a slot
UI-Venus-2, released on 2026-08-27, is a multimodal GUI agent that finally bridges the gap between benchmark performance and real-world reliability. It's not just another model that aces tests but fails in production—this one is built to handle the messiness of actual digital environments. The key is its unified reasoning-action loop, which expands environment coverage to over 170 multilingual mobile apps and native desktop operating systems. That's a big deal because most GUI agents fall apart when they encounter an app or OS they weren't trained on. UI-Venus-2 also uses a deep-research pipeline for function-grounded instruction generation, meaning it understands tasks at a functional level rather than just matching surface patterns. But the real innovation is in verification. UI-Venus-2 uses trace-level and sample-level evaluators with visual keypoints and multi-model voting to ensure its actions are correct. This isn't just about avoiding mistakes—it's about providing reliable reinforcement learning signals during training. The agent also integrates safety mechanisms to control consequential actions, which is non-negotiable for any system operating in the real world. I think the open-source release is the most important part. GUI agents have been stuck in research labs because transitioning to practical deployment is so hard. By making UI-Venus-2 available, the Venus Team is giving developers a foundation they can actually build on. We don't know yet if this will become the standard for real-world digital automation, but it's the first candidate on this desk that doesn't arrive with an asterisk. The paper is at [UI-Venus-2 Technical Report](https://arxiv.org/abs/2609.00028). The project page is at [https://flesymeb.github.io/UI-Venus-2/](https://flesymeb.github.io/UI-Venus-2/).