Florence-2 Large
Free.ai (self-hosted)
·
ocr
·
~300 tokens per page
Florence-2 Large is an OCR model built by Microsoft. Strongest at Complex document understanding, region-grounded OCR.. Self-hosted on Free.ai GPUs — runs free against your daily token pool (300 tokens per page). Released under MIT — commercial use permitted on Free.ai.
Use via API
OpenAI-compatible REST API. Generate a key and call this model in seconds.
curl -X POST https://api.free.ai/v1/ocr/ \
-H "Authorization: Bearer sk-free-..." \
-H "Content-Type: application/json" \
-d '{"model":"florence-2","prompt":"your prompt here"}'Frequently Asked Questions
Florence-2 Large extracts text from images and PDFs. Upload a scanned page, a screenshot, or a photo of a document and Florence-2 Large returns the text — preserving line breaks and (on capable engines) tables, formulas, and structure.
Florence-2 Large handles dozens of scripts — Latin, Cyrillic, CJK, Arabic, Devanagari, and more. The /ocr/ landing has the full language list per engine.
Yes for clean printed-style handwriting; cursive and shorthand are harder. TrOCR and modern transformer-based engines are noticeably better than legacy Tesseract on handwriting.
Layout-aware engines (GOT-OCR, Marker, Florence-2) return Markdown/HTML tables and LaTeX for formulas. Plain-text engines return text without structure — check this model's output mode.
Florence-2 Large runs on our GPUs from your daily free pool; $5 → 200,000 paid tokens after. ~300 tokens per page.
PNG, JPG, WebP, HEIC, BMP, plus single- and multi-page PDFs up to 50 MB. Skewed or rotated pages are auto-corrected before extraction.
Character accuracy is typically >98% on clean printed pages, dropping on low-res photos, severely skewed scans, or unusual fonts. Compare engines on /ocr/compare/ when accuracy matters.
Yes — /batch/ accepts a folder of images/PDFs. Each extraction lands in /account/?tab=history. For folder-tree preservation use the API.
Yes — POST your file to /v1/ocr/ with model="Florence-2 Large". Returns JSON with text + (when supported) layout + word-level bounding boxes. /api/ has the full reference.
Self-hosted models keep your file on our GPUs; premium pass through with a DPA. Documents are deleted after the share-window. We do not train on your inputs.
Yes — Free.ai grants commercial use of extracted text. You need rights to the source document.
1–5 seconds per page typically. Layout-aware engines (tables, math) take longer. Multi-page PDFs scale roughly linearly.