GLADOS-1 — UI-TARS-7B-SFT

Model Description

GLADOS-1 is the first computer-use (CUA) model post-trained using collective, crowd-sourced trajectories. Leveraging the enourmous PANGO dataset (with primarily Chrome based interactions), it's purpose is to provide a lense as to what's possible with enormous trajectory sizes in computer use.

It also represents the first open-sourced post-training pipeline for UI-TARS, inspired by the existing Qwen2VL finetuning series.

This model is designed to:

• Be compliant. It has been taught to rigorouly follow directions and output action formats compatible with downstream parsers like PyAutoGUI.
• Understand web productivity applications. The Pango dataset primarily contains productivity application usage in browser. Consequently in OSWorld results, we observe significantly improved performance on the Chrome task bench.
• Have strong intuition on visual grounding. Our experiments are detailed more closely here in our research blog.

📕 Release Blog   |    🤗 Code   |    🔧 Deployment (via UI-TARS)    |    🖥️ Running on your own computer (via UI-TARS Desktop)  

Citation

@misc{chakralabs2025glados-1,
  author = {Chakra Labs},
  title = {GLADOS-1},
  url = {https://github.com/Chakra-Network/GLADOS-1},
  year = {2025}
}