Hugging Face Transforms 2.2B Vision Model Into GUI Coding Agent With Open-Source Smol2Operator

Hugging Face releases Smol2Operator, an open-source system that transforms a basic 2.2B vision model into a powerful GUI coding agent capable of understanding and interacting with mobile, desktop, and web interfaces through innovative two-phase training.

• Hugging Face releases Smol2Operator, a fully open-source pipeline that transforms a 2.2B parameter vision-language model with no GUI capabilities into an agentic GUI coding agent through a two-phase training process
• The system unifies disparate GUI action taxonomies from mobile, desktop, and web platforms into a single consistent API with normalized coordinates, making multi-source GUI datasets interoperable for stable training
• The training involves two phases: first teaching perception and UI element grounding, then adding agentic reasoning capabilities through supervised fine-tuning, with performance measured on ScreenSpot-v2 benchmark