SmolDocling: An Ultra-Compact Vision-Language Model for End-to-End Multi-Modal Document Conversion (arXiv, March 2025)

Forum topic · 小凯 · 2026-07-05

Summary

SmolDocling is a 256M-parameter vision-language model introduced in a March 2025 arXiv paper (arXiv:2503.11576) for end-to-end conversion of document page images into structured markup. Unlike multi-stage pipelines that separately run OCR, layout detection, table recognition, and equation extraction, SmolDocling performs the entire conversion in a single model. It outputs DocTags, a compact and machine-readable document representation that explicitly captures the spatial relationship between text blocks and page structure, which can be losslessly transformed into common formats such as Markdown and HTML. According to the paper, the model targets accurate recognition of core document elements including text, tables, equations, charts, code, and section headers while reducing model size and computational cost relative to larger general-purpose VLMs. The authors release the model weights and benchmark datasets to the community, positioning SmolDocling as an efficient open baseline for document understanding and conversion tasks.

SmolDocling: An Ultra-Compact Vision-Language Model for End-to-End Multi-Modal Document Conversion

This post introduces an academic paper entry: SmolDocling: An ultra-compact vision-language model for end-to-end multi-modal document conversion (March 2025, arXiv).

Key information

| Field | Content | |---|---| | Title | SmolDocling: An ultra-compact vision-language model for end-to-end multi-modal document conversion | | Authors / Affiliations | Ahmed Nassar, Andres Marafioti, Matteo Omenetti, Maksym Lysak, Nikolaos Livathinos, Christoph Auer, et al. (13 authors total) | | Published | March 2025 | | Source | https://arxiv.org/abs/2503.11576 | | Type | Academic paper | | Category | Document understanding |

One-line summary

SmolDocling is an ultra-compact vision-language model that performs end-to-end conversion of multi-modal documents into structured, machine-readable markup.

Original abstract (preserved)

> SmolDocling: An ultra-compact vision-language model for end-to-end multi-modal document conversion, Mar 2025, arxiv

Context and relevance

The paper is positioned in the document understanding area, addressing how to replace conventional multi-stage document conversion pipelines (OCR, layout detection, table and equation recognition) with a single compact vision-language model. Related entries in the collection include:

LongDA: Benchmarking LLM Agents for Long-Document Data Analysis
Qwen2.5-VL Technical Report (Section 3.3.2: Document Understanding and OCR)
ColPali: Efficient Document Retrieval with Vision Language Models (https://arxiv.org/abs/2407.01449)

Notes for readers

Full quantitative results should be verified against the original PDF on arXiv before being cited.
The work is cross-indexed with other entries along the retrieval → ranking → generation/agent → evaluation chain in this collection.

Reference

Original paper: SmolDocling: An ultra-compact vision-language model for end-to-end multi-modal document conversion, arXiv, March 2025. https://arxiv.org/abs/2503.11576