hugging-apps/deskforge-gui-grounding
DeskForge · GUI Grounding for Computer-Use Agents
This demo runs docling-project/DeskForge-Qwen3.5-4B, a GUI-grounding computer-use agent fine-tuned from Qwen3.5-4B on the DeskForge-1M corpus (1.21M annotated desktop observations, 917K recorded click transitions).
Give it a desktop screenshot plus a task, and it replies with the next action as pyautogui code. Coordinates are fractions of the screen: x runs from 0.0 at the left edge to 1.0 at the right edge, y from 0.0 at the top to 1.0 at the bottom. The predicted click point is drawn on the screenshot.
- Paper: DeskForge: Dense Supervision from Desktop Environments for Computer-Use Agents
- Project page: saidgurbuz.github.io/deskforge
- Code: github.com/Saidgurbuz/deskforge
The bundled example screenshots are held-out validation transitions from the DeskForge-1M dataset (MIT license, demo/transitions_preview subset).
Note: Qwen3.5's hybrid linear-attention layers run their reference PyTorch kernels on this hardware (the optimized causal_conv1d / flash-linear-attention extensions are not installed), which is correct but slower than the fused kernels.
