docling-project/DeskForge-Qwen3.5-4B

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 30, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

DeskForge-Qwen3.5-4B by docling-project is a 4.5 billion parameter vision-language model, fine-tuned from Qwen3.5-4B with a 32768 token context length. It specializes in desktop GUI grounding and next-action prediction, translating natural language instructions and screenshots into `pyautogui` commands. This model excels at automating desktop interactions, achieving significant improvements on external GUI benchmarks and long-horizon task completion compared to its base model.

Loading preview...

DeskForge-Qwen3.5-4B Overview

DeskForge-Qwen3.5-4B is a 4.5 billion parameter model developed by docling-project, fine-tuned from the Qwen3.5-4B base model. Its primary function is to interpret desktop screenshots and natural language instructions to predict the next GUI action as a pyautogui call, including precise fractional coordinates. This model is specifically designed for desktop GUI grounding and next-action prediction, leveraging extensive GUI-workflow supervision during its fine-tuning process.

Key Capabilities

  • Desktop GUI Automation: Translates user instructions and visual input (screenshots) into executable pyautogui commands for desktop interaction.
  • Enhanced Grounding Performance: Achieves an 81.56% mean accuracy on DeskForge-1M held-out conditions and significantly outperforms its base model on external GUI benchmarks like ScreenSpot-Pro (49.02% vs 30.17%) and MMBench-GUI (80.13% vs 61.83%).
  • Long-Horizon Task Completion: Demonstrates improved performance in multi-step tasks, solving 50 tasks on WebArena-Infinity and 15 on OpenApps with its v1.0 checkpoint, compared to 31 and 3 tasks respectively for the base Qwen3.5-4B.
  • Vision-Language Integration: Processes visual information from screenshots (capped at 2,097,152 pixels) to inform action predictions.

Good For

  • Developing Computer-Use Agents: Serves as a robust action model for agents designed to automate desktop workflows and interact with graphical user interfaces.
  • GUI Testing and Automation: Ideal for scenarios requiring precise, instruction-driven control over desktop applications.
  • Research in Human-Computer Interaction: Provides a strong foundation for exploring advanced methods of desktop interaction and task automation.

Finetunes
1 models