Chandra vs SmolDocling
Pick Chandra for production conversion of multilingual or scanned documents, where it outputs Markdown, HTML or JSON with bounding boxes and has a hosted API option. Pick SmolDocling only for English experiments on small hardware: it is a 256M preview model, and its authors point users to Granite-Docling for updates and support. The main split is maturity and language coverage: Chandra scores 77.8% on a 43-language benchmark, while SmolDocling is English only.
Chandradatalab/chandra | SmolDoclinghuggingface/smoldocling | |
|---|---|---|
| Type | Open source | Open source |
| Price | 2 credits / page | 1 credit / page |
| Per 1,000 pages | $2.00 | $1.00 |
| Typical speed | 2.0 s / page | 1.0 s / page |
| Best for | Complex layouts and handwriting | Very cheap page conversion |
| Inputs | PDF, images | PDF, images |
| Languages | 40+ languages | Mainly English |
| Max pages per request | 500 | 500 |
| Tables | ✓ Yes | ✓ Yes |
| Handwriting | ✓ Yes | No |
| Math | ✓ Yes | ✓ Yes |
| Bounding boxes | ✓ Yes | ✓ Yes |
| Confidence | No | No |
| Forms | ✓ Yes | No |
| Multilingual | ✓ Yes | No |
| On OCRRouter | Available | Available |
Prices are what you pay on OCRRouter. Speed is the provider's typical figure, not a benchmark we ran.
Which to choose
Choose Chandra if
- Your documents are not only English: Chandra 2 averages 77.8% across 43 languages, and 72.7% on a 90-language benchmark.
- You want a hosted option as well as open weights: the Datalab API costs $1, $3 or $7.50 per 1,000 pages, with a $20/month free tier.
- You need Markdown, HTML or JSON with bounding boxes and labels per block, for RAG pipelines or documents with tables, forms and math.
Choose SmolDocling if
- You want the smallest model: 256M parameters, small enough for modest hardware.
- You need a free licence with no revenue threshold: Apache-2.0 per the model summary, self-hosted only.
- You want speed on English pages: about 0.35 seconds per page on an A100 with vLLM.
Key differences
| Chandra | SmolDocling | |
|---|---|---|
| Licence | Code Apache-2.0; weights modified OpenRAIL-M, free for research, personal use and companies under $2M funding/revenue | Apache-2.0 per model summary (metadata tag says CDLA-Permissive-2.0) |
| How you run it | Self-host via vLLM or Hugging Face, or use the hosted Datalab API | Self-host only |
| Price | Self-hosted: free within licence terms. API: $1, $3 or $7.50 per 1,000 pages | Free, self-hosted (you pay for compute) |
| Languages | 43-language average 77.8%; weak on Urdu (43.2%) and Sanskrit (51.1%) | English only |
| Output formats | Markdown, HTML, JSON, bounding boxes | DocTags, Markdown, HTML, JSON, LaTeX, bounding boxes |
| Benchmark | olmOCR-bench overall 85.8 (Chandra 2, authors' table) | Table TEDS 0.82 on FinTabNet, full-page OCR F1 0.80; below Granite-Docling on every reported task |
| Known failure | Token-loop hallucinations on 5-6 of about 60 Russian/Kazakh scans in one issue report; one reviewer measured 66 seconds per page on an RTX 4090 with the Hugging Face path | Endless token repetition and missing closing tags; output loops on some images, making generation very slow even on sparse pages |
Sources
- Chandra GitHub README
- Datalab pricing
- GitHub issue #51 token loops
- SmolDocling model card
- Docling issue #1515
- Granite-Docling comparison table
What each one is
Chandra
Datalab's OCR model that keeps layout and handles forms, tables and handwriting.
Chandra detailsSmolDocling
Tiny 256M vision model from Hugging Face and IBM that outputs DocTags. Predecessor of Granite Docling.
SmolDocling detailsTry both on your own file
Run the file once with each model and compare the output.
Use one, fall back to the other
{"model": "datalab/chandra","fallbacks": ["huggingface/smoldocling"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
Chandra alternativesSmolDocling alternativesBest OCR models in 2026