SmolDocling
256M vision-language model for end-to-end document conversion.
About
SmolDocling is a 256M-parameter vision-language model from IBM Research and Hugging Face that converts whole document pages — tables, equations, charts, code listings, forms — into DocTags, a markup format that preserves content, structure, and spatial location in one pass. The paper reports it competing with VLMs up to 27 times its size while running on a fraction of the compute.
At 256M parameters it is a specialist, not a generalist: unusual layouts and low-quality scans still challenge it, and the preview release predates the final datasets. Confirm the license on the Hugging Face model page before commercial use.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Competes with VLMs up to 27x larger on full-page document conversion (paper-reported, arXiv 2503.11576)
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (metrics tier, metrics-v1). Evidence: IBM Research x Hugging Face, Mar 2025; paper (arXiv 2503.11576) plus preview weights and demo Space; 256M specialist for document conversion; DocTags markup format is a novel contribution; Preview release — final datasets were promised but not confirmed here; confirm license on the model card; Benchmark claim (competes with 27x larger VLMs) is paper-reported, not independently verified. License: Apache-2.0 per the Hugging Face model page — re-confirm before commercial use.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 6.0/10
- Maintenance · recency, releases, issue load
- 6.0/10
- Openness · license + docs
- 8.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · IBM Research x Hugging Face, Mar 2025; paper (arXiv 2503.11576) plus preview weights and demo Space
- · 256M specialist for document conversion; DocTags markup format is a novel contribution
- · Preview release — final datasets were promised but not confirmed here; confirm license on the model card
- · Benchmark claim (competes with 27x larger VLMs) is paper-reported, not independently verified