cs.AIAug 27, 2026
SaveUI-Venus-2 Technical Report
Organizations: Ant Group
Abstract
Multimodal GUI agents have emerged as a promising paradigm for digital task automation, yet transitioning from benchmark-oriented models to dependable real-world applications remains challenging due to limited environment coverage, brittle task construction, and unreliable reward verification. In this work, we present UI-Venus-2, a general-purpose foundation GUI agent designed to operate across mobile, web, and desktop environments through a unified closed-loop reasoning-action framework. To bridge the gap toward practical deployment, we jointly scale three critical dimensions: (1) Environments, expanding coverage to more than 170 multilingual mobile apps and native desktop operating systems; (2) Tasks, employing a deep-research pipeline for function-grounded instruction generation; and (3) Verification, adopting trace-level and sample-level evaluators with visual keypoints and multi-model voting to ensure reliable RL signals for training. Furthermore, we integrate safety-aware mechanisms to ensure controlled execution of consequential actions. By offering a capable, efficient, and open-source foundation, UI-Venus-2 advances the field toward more generalizable, verifiable, and self-reflective agents for real-world applications.
Figures & tables
Figure 1: 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 the results are reproduced by us.
Figure 2: System Overview of UI-Venus-2 . The figure illustrates the task generation and trajectory collection process of UI-Venus-2. Diverse tasks are constructed to form a multi-domain task pool, and interaction trajectories are collected across mobile, browser, and computer environments, covering a broad range of real-world applications, websites, and desktop software.
Figure 3: The Three-Stage Pipeline of UI-Venus-2 . Following the overall training recipe of UI-Venus-1.5, UI-Venus-2 starts with large-scale trajectory-based mid-training to inject GUI interaction knowledge. The resulting model is then optimized independently for each domain using step-level Offline-RL, covering Grounding, CAPTCHA, Mobile, Web, and Computer tasks. Finally, the domain-specialized models are consolidated into the final UI-Venus-2 model via multi-teacher on-policy distillation.
| Models | MobileGym | VenusBench-Mobile | AndroidWorld | MobileWorld | KnowUBench | MemGUI |
|---|---|---|---|---|---|---|
| General VLMs | ||||||
| Qwen3.5-9B ( Qwen Team, 2026a ) | 9.0* | 15.3* | 57.8 | 18.0(18.0)* | 33.3 | 6.2* |
| Qwen3.6-27B ( Qwen Team, 2026b ) | 24.6* | 28.0* | 70.3 | 36.8(41.9)* | - | 25.7* |
| Claude-Opus-4.6 ( Anthropic, 2026a ) | - | 36.5* | - | 44.5 | - | - |
| Kimi-K2.6 ( Moonshot AI, 2026a ) | 38.7* | 31.2* | - | 55.6 | - | 39.1 |
| Kimi-K3 ( Moonshot AI, 2026b ) | - | - | - | 74.4 | - | - |
Table 1: 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 the baseline results evaluated or reproduced by us.
| Models | OSWorld-Verified | DeskCraft |
|---|---|---|
| General VLMs | ||
| Claude-Opus-4.8 ( Zhou et al., 2026 ) | 83.4 | - |
| Qwen3.5-9B ( Qwen Team, 2026a ) | 41.8 | 14.6 ∗ |
| Qwen3.6-27B ( Qwen Team, 2026b ) | 62.0 | 28.7 ∗ |
| Kimi-K2.6 ( Moonshot AI, 2026a ) | 73.1 | 41.4 ∗ |
| Seed-2.0-Pro ( Seed, 2026 ) | 62.3 | 40.0 ∗ |
Table 2: 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.
| Models | WebVoyager | Online-Mind2Web | REAL | Odysseys | |
|---|---|---|---|---|---|
| Avg. | Perfect | ||||
| General VLMs | |||||
| Qwen3.5-9B ( Qwen Team, 2026a ) | 46.9 ∗ | 27.3 ∗ | 18.2 ∗ | 42.6 ∗ | 13.5 ∗ |
| Qwen3.5-4B ( Qwen Team, 2026a ) | – | – | – | 42.9 | 10.7 |
| Qwen3.6-27B ( Qwen Team, 2026b ) | 84.3 ∗ | 55.3 ∗ | 27.3 ∗ | 39.5 ∗ | 18.5 ∗ |
| OpenAI Operator ( OpenAI, 2025 ) | 87.0 | 61.3 | – | – | – |
Table 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.
| Models | Grounding Benchmarks | |||
|---|---|---|---|---|
| VenusBench-GD | ScreenSpot-Pro | OSworld-G-R | UI-Vision | |
| General VLMs | ||||
| Qwen 3.7 Plus ( Qwen Team, 2026c ) | 75.2* | 68.9 | 78.2 | 68.0 |
| Seed 2.1 Pro ( ByteDance Seed, 2026 ) | 73.9* | 65.3 | 78.0 | 62.0 |
| Kimi-K2.6 ( Moonshot AI, 2026a ) | 73.1* | 52.0* | 69.7* | 51.7* |
| Qwen3.6-27B ( Qwen Team, 2026b ) | 67.7* | 65.2* | 76.9* | 58.3* |
Table 4: Performance comparison on various Grounding Benchmarks . VenusBench-GD reports English-instruction micro-average point-in-box accuracy. “*” indicates baselines evaluated or reproduced by us.
| Models | VenusBench- CAPTCHA | MCA-Bench | Spatial- CAPTCHA-Bench | NextGen- CAPTCHAs | Open CaptchaWorld |
|---|---|---|---|---|---|
| General VLMs | |||||
| Qwen3.5-9B ( Qwen Team, 2026a ) | 28.3 | 30.4 | 4.9 | 2.8 | 36.4 |
| Qwen3.6-27B ( Qwen Team, 2026b ) | 53.0 | 51.7 | 31.0 | 14.1 | 47.7 |
| Seed-2.0-Pro ( Seed, 2026 ) | 47.9 | 36.5 | 43.8 | 20.4 | 55.6 |
| Kimi-K2.6 ( Moonshot AI, 2026a ) | 39.7 | 38.7 | 24.8 | 7.2 | 47.8 |
| Claude-Opus-4.6 ( Anthropic, 2026a ) | 16.0 | 25.9 | 9.5 | 2.8 | 23.3 |
Table 5: 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 Wu et al. (2026b) , Spatial-CAPTCHA-Bench Kharlamova et al. (2026) , NextGen-CAPTCHAs Liu et al. (2026b) , and Open CaptchaWorld Luo et al. (2025b) . We use 1,000 sampled MCA-Bench examples, 15 NextGen-CAPTCHAs task types, and 16 Open CaptchaWorld task types; see the appendix for selection details.
| Model | OSHarm (ASR ) | OSBlind (ASR ) |
|---|---|---|
| General VLMs | ||
| Qwen3.5-9B ( Qwen Team, 2026a ) | 25.3 | 79.4 |
| Qwen3.5-27B ( Qwen Team, 2026a ) | 18.0 | 89.3 |
| Kimi K2.6 ( Moonshot AI, 2026a ) | 32.0 | 93.6 |
| GUI-specific Models | ||
| EvoCUA-8B ( Xue et al., 2026 ) | 39.3 | 85.3 |
Table 6: Safety evaluation on OSHARM and OS-BLIND benchmarks. ASR denotes Attack Success Rate (%, lower is better).
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
| Action | Definition |
|---|---|
| Shared Actions (All Platforms) | |
| Click(point=(x, y)) | Click at coordinates (x, y). |
| Drag(start=(x1, y1), end=(x2, y2)) | Drag from (x1, y1) to (x2, y2). |
| Swipe(start=(x1, y1), end=(x2, y2)) | Scroll by swiping from (x1, y1) to (x2, y2). |
| DoubleClick(point=(x, y)) | Perform a double-click/tap at coordinates (x, y). |
| LongPress(point=(x, y)) | Long press at coordinates (x, y) to trigger extra options. |
Table 7: All actions and their definitions used in UI-Venus-2 . We unify the action space and map all the actions in the existing open-source dataset to this space.
Figure 4: Examples of synthesized GUI Grounding training data. Our pipeline generates diverse and realistic interface screenshots spanning desktop (macOS, Windows), mobile (iOS), and web platforms, covering both professional software and consumer applications. All interfaces are rendered in a real headless Chromium browser via Playwright, ensuring high-fidelity visual output that closely mirrors authentic user environments.
Figure 5: Overview of VenusBench-CAPTCHA. Each panel shows one complete, uncropped representative screenshot. The OCR label gives the target transcription, numbered boxes indicate the required click order, and arrows visualize annotated drag trajectories. These annotations are added for presentation only and are not part of the model input.
More in this topic
GUI Agents
Graphical user interface (GUI) agents perceive and operate apps through screens, controls, and other software interfaces.
See all 178 papers