---
title: DeskForge Gemma4 E4B GUI Grounding
emoji: 🖱️
colorFrom: indigo
colorTo: blue
sdk: gradio
sdk_version: 6.30.0
app_file: app.py
short_description: Ground desktop UI elements from a screenshot and task.
python_version: "3.12"
startup_duration_timeout: 30m
license: apache-2.0
---

# DeskForge-Gemma4-E4B · GUI Grounding

This demo runs [docling-project/DeskForge-Gemma4-E4B](https://huggingface.co/docling-project/DeskForge-Gemma4-E4B),
a GUI-grounding computer-use action model fine-tuned from
[Gemma 4 E4B](https://huggingface.co/google/gemma-4-E4B-it) on 200K grounding
examples from the [DeskForge-1M](https://huggingface.co/datasets/docling-project/DeskForge-1M)
desktop corpus.

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](https://arxiv.org/abs/2610.02320)
- Project page: [saidgurbuz.github.io/deskforge](https://saidgurbuz.github.io/deskforge/)
- Code: [github.com/Saidgurbuz/deskforge](https://github.com/Saidgurbuz/deskforge)

The bundled example screenshots are held-out validation transitions from the
DeskForge-1M dataset (MIT license).

**Note:** pinned to `transformers==5.16.1` — the version the model card tests and
the one matching the checkpoint's `transformers_version`. From 5.18.0 onward the
Gemma4 vision tower requires `rope_type: "axial"`, which this checkpoint's
`vision_config` (shipped as `rope_type: "default"`) does not declare.