SmolDocling: A 256M Open-Source VLM

Researchers from IBM and Hugging Face have recently addressed these challenges by releasing SmolDocling, a 256M open-source vision-language model (VLM) designed explicitly for end-to-end multi-modal document conversion tasks. Unlike larger foundational models, SmolDocling provides a streamlined solution that processes entire pages through a single model, significantly reducing complexity and computational demands. Its ultra-compact nature, at just 256 million parameters, makes it notably lightweight and resource-efficient. The researchers also developed a universal markup format called DocTags, which precisely captures page elements, their structures, and spatial contexts in a highly compact and clear form.

https://huggingface.co/ds4sd/SmolDocling-256M-preview

https://arxiv.org/pdf/2503.11576

3 comments