dots.ocr

dots.ocr is the open-weight multilingual document OCR model from rednote-hilab (Rednote / Xiaohongshu’s research lab), released MIT in July 2025. A compact 3B model built around true multilingual coverage — 100+ languages including Devanagari, which makes it the natural pick for documents in Nepali, Hindi, Bengali, or the long tail of non-Latin scripts.

We host it on our GPUs as the multilingual specialist in the OCR catalog — between GLM-OCR’s Chinese + English baseline and olmOCR-2’s English / academic premium tier.

When to pick dots.ocr

• Nepali, Hindi, or other Devanagari-script content — government documents, legal records, regional publications, news articles
• Multilingual document pipelines where the input script varies
• Mixed-script documents — an English report with Nepali tables, a Hindi document with English citations
• Indic / Arabic / Thai / SEA-script workloads the Western-trained models miss

When to pick something else

• English-only academic or legal → olmOCR-2 is trained specifically for that
• High-volume Chinese / English enterprise paperwork → GLM-OCR is the cheapest and most accurate for that workload

Pricing

EUR 0.0020 per page. Flat. No per-token tail. Pages detected automatically for PDFs; a single image counts as one page.

Limits

• Per-tenant rate limit: 30 pages per second
• Image size limit: 50 MB per page (PDFs split automatically)
• Supported formats: png, jpg, webp, pdf, tiff

Output formats

Raw text, markdown with preserved structure (HTML tables, LaTeX equations), JSON with caller-supplied schema, and interleaved text + bounding boxes with grounded coordinates for audit pipelines.

Why we run it

A Nepal-based AI cloud should have first-class Nepali document support. dots.ocr is currently the strongest open-weight option for Devanagari accuracy at modern parser quality — we host it because the customer base needs it.