Model Description
Qwen3-VL-32B-Thinking-Mobile-Teacher is fine-tuned from Qwen3-VL-32B-Thinking on mobile interaction trajectories from the Uni-GUI dataset. It serves as the mobile-platform teacher in the UI-MOPD multi-teacher on-policy distillation framework.
Key Highlights
- Base Model: Qwen3-VL-32B-Thinking
- Training Data: Mobile subset of Uni-GUI (~160K interaction steps across ~11.5K trajectories)
- Role: Platform-specific teacher for mobile environments in the UI-MOPD distillation pipeline
- MobileWorld Performance: 16.2% task success rate (vs. 9.4% base model)
Training Details
This model is obtained in Stage 1 of the UI-MOPD training pipeline:
- Stage 1 (This Model): Supervised fine-tuning of Qwen3-VL-32B-Thinking on mobile GUI interaction trajectories from Uni-GUI to produce a platform-specific mobile expert.
- Stage 2: The mobile teacher (this model) and a desktop teacher jointly guide a shared 8B student policy via multi-teacher on-policy distillation with platform-conditioned routing.
Table with columns: Method, MobileWorld| Method | MobileWorld |
|---|
| Qwen3-VL-32B-Thinking (base) | 9.4% |
| Mobile Teacher (this model) | 16.2% |
Intended Use
This model is designed to:
- Serve as a teacher model in the UI-MOPD distillation framework for training cross-platform GUI agents
- Be used as a standalone mobile GUI agent for executing mobile tasks (e.g., app navigation, settings control, messaging)
How to Use
from transformers import Qwen3VLForConditionalGeneration, AutoProcessor
model = Qwen3VLForConditionalGeneration.from_pretrained(
"UI-MOPD/Qwen3-VL-32B-Thinking-Mobile-Teacher",
torch_dtype="auto",
device_map="auto",
)
processor = AutoProcessor.from_pretrained("UI-MOPD/Qwen3-VL-32B-Thinking-Mobile-Teacher")
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},
}