UI-MOPD/Qwen3-VL-8B-Thinking-UI-MOPD-Student

VISIONConcurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 4, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

UI-MOPD/Qwen3-VL-8B-Thinking-UI-MOPD-Student is an 8 billion parameter multimodal GUI agent model developed by UI-MOPD, based on Qwen3-VL-8B-Thinking. It is trained using multi-teacher on-policy distillation to achieve balanced cross-platform performance for GUI automation on both desktop and mobile environments. This model excels at interactive task execution and preserves GUI grounding capabilities, making it suitable for cross-platform GUI agent applications.

Loading preview...

Overview

UI-MOPD/Qwen3-VL-8B-Thinking-UI-MOPD-Student is an 8 billion parameter multimodal GUI agent, distilled from the Qwen3-VL-8B-Thinking base model. Developed by UI-MOPD, this "Student" model is designed for continual GUI agent learning across different platforms.

Key Capabilities

  • Cross-Platform GUI Automation: Achieves balanced performance on both desktop (OSWorld) and mobile (MobileWorld) environments.
  • Multi-Teacher Distillation: Learns from two 32B platform-specific teachers (Desktop and Mobile) using reinforcement learning with platform-conditioned distillation.
  • Enhanced Task Success: Demonstrates a 38.2% task success rate on OSWorld and 12.0% on MobileWorld, showing significant improvement over the base model.
  • Preserves Visual Understanding: Maintains strong GUI grounding and visual understanding capabilities, as evidenced by performance on benchmarks like AndroidControl and OSWorld-G.

Training Methodology

This model is the result of Stage 2 of the UI-MOPD training pipeline, which involves:

  • Reinforcement Learning (DAPO): Utilizes Deep Actor-Critic Policy Optimization.
  • Platform-Conditioned Routing: Dynamically selects the appropriate platform-specific teacher during training.
  • Adaptive KL Masking: Prevents over-regularization by removing teacher penalties when task rewards are sufficient.

Intended Use Cases

  • Developing cross-platform GUI agents for automating tasks on desktop applications (e.g., web browsing, file management) and mobile apps (e.g., navigation, settings control).
  • Research into continual learning and multi-platform adaptation for GUI agents.

Finetunes
1 models