Model Description
Qwen3-VL-8B-Thinking-UI-MOPD-Student is trained from Qwen3-VL-8B-Thinking using the UI-MOPD framework. It learns from two platform-specific 32B teachers (Desktop Teacher and Mobile Teacher) through reinforcement learning with platform-conditioned distillation, achieving balanced cross-platform performance on both desktop and mobile environments.
Key Highlights
- Base Model: Qwen3-VL-8B-Thinking
- Training Method: Multi-teacher on-policy distillation with DAPO + platform-conditioned KL regularization
- Teachers: Qwen3-VL-32B-Thinking-Desktop-Teacher + Qwen3-VL-32B-Thinking-Mobile-Teacher
- OSWorld Performance: 38.2% task success rate
- MobileWorld Performance: 12.0% task success rate
Training Details
This model is obtained in Stage 2 of the UI-MOPD training pipeline:
- Stage 1: Supervised fine-tuning of Qwen3-VL-32B-Thinking on platform-specific data to produce a Desktop Teacher (46.3% on OSWorld) and a Mobile Teacher (16.2% on MobileWorld).
- Stage 2 (This Model): The 8B student is trained with reinforcement learning (DAPO) combined with multi-teacher on-policy distillation. A platform-conditioned router selects the appropriate teacher based on the current environment, and adaptive KL masking prevents over-regularization.
Key Training Components
- Platform-Conditioned Routing: Routes each rollout to the corresponding platform-specific teacher
- K3 Estimator: Efficient single-sample KL divergence estimator
- Adaptive KL Masking: Removes teacher penalty when task reward is already sufficient
Table with columns: Method, OSWorld (Desktop), MobileWorld (Mobile)| Method | OSWorld (Desktop) | MobileWorld (Mobile) |
|---|
| Qwen3-VL-8B-Thinking (base) | 33.9% | 7.7% |
| Mixed-SFT | 35.0% | 6.4% |
| Model Merge (TIES) | 36.8% | 0% |
| UI-MOPD (this model) | 38.2% | 12.0% |
UI-MOPD achieves state-of-the-art balanced cross-platform performance, with +12.7% relative improvement on OSWorld and +55.8% on MobileWorld compared to the base model.
GUI Grounding & Understanding
Table with columns: Model, AndroidControl, ScreenSpot-Pro, ScreenSpotV2, OSWorld-G| Model | AndroidControl | ScreenSpot-Pro | ScreenSpotV2 | OSWorld-G |
|---|
| Qwen3-VL-8B-Thinking (base) | 78.73% | 43.71% | 91.27% | 52.13% |
| Model Merge (TIES) | 74.01% | 37.13% | 88.60% | 47.16% |
| UI-MOPD (this model) | 80.05% | 43.14% | 90.88% | |
UI-MOPD preserves GUI grounding and visual understanding capabilities while improving interactive task performance.
Intended Use
- Cross-platform GUI agent for executing tasks on both desktop (e.g., web browsing, file management) and mobile (e.g., app navigation, settings control) environments
- Research on continual learning and multi-platform adaptation for GUI agents
How to Use
from transformers import Qwen3VLForConditionalGeneration, AutoProcessor
model = Qwen3VLForConditionalGeneration.from_pretrained(
"UI-MOPD/Qwen3-VL-8B-Thinking-UI-MOPD-Student",
torch_dtype="auto",
device_map="auto",
)
processor = AutoProcessor.from_pretrained("UI-MOPD/Qwen3-VL-8B-Thinking-UI-MOPD-Student")
Citation
@misc{lian2026uimopdmultiplatformonpolicydistillation,
title={UI-MOPD: Multi-Platform On-Policy Distillation for Continual GUI Agent Learning},
author={Niu Lian and Alan Chen and Zhehao Yu and Chengzhen Duan and Fazhan Liu and Hui Liu and Pei Fu and Jian Luan and Yaowei Wang and Shu-Tao Xia and Jinpeng Wang},
year={2026},
eprint={2607.04425},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2607.04425},
}