UI-MOPD/Qwen3-VL-8B-Thinking-Mobile-SFT

Model Overview

UI-MOPD/Qwen3-VL-8B-Thinking-Mobile-SFT is an 8 billion parameter multimodal model, supervised fine-tuned (SFT) from the Qwen3-VL-8B-Thinking base model. Developed by UI-MOPD, this model is specifically trained on mobile GUI interaction data from the Uni-GUI dataset, encompassing approximately 160,000 interaction steps across 11,500 trajectories.

Key Capabilities & Features

• Mobile GUI Agent: Functions as a standalone agent for executing tasks within mobile graphical user interfaces, such as app navigation, settings control, and messaging.
• Multimodal: Leverages the vision-language capabilities of its Qwen3-VL-8B-Thinking base.
• Data Quality Reference: Released as an auxiliary artifact to demonstrate the effectiveness and quality of the Uni-GUI SFT data at the 8B parameter scale.
• Enables Model-Merge Experiments: Provides a suitable base for community research involving model merging techniques (e.g., TIES, DARE) with other 8B GUI agents.

Relationship to UI-MOPD Pipeline

It's important to note that this specific SFT model is not part of the primary UI-MOPD student training pipeline. The UI-MOPD student model is cold-started directly from the Qwen3-VL-8B-Thinking base and trained via on-policy distillation from 32B teacher models. This Mobile-SFT model is an independently trained checkpoint, offering a baseline for SFT-only performance compared to distillation-enhanced approaches.

Intended Use Cases

• Developing and deploying mobile GUI automation solutions.
• Researching and evaluating the impact of Uni-GUI SFT data.
• Experimenting with model merging strategies for GUI agents.
• Establishing a performance baseline for SFT-only mobile GUI models.