What is Falcon OCR?
Falcon OCR is an open document-OCR model from the Technology Innovation Institute in Abu Dhabi, released under Apache 2.0 and described in the Falcon Perception technical report of 28 March 2026. It turns a page image into plain text, LaTeX for formulas or HTML for tables. Version 1.5, uploaded on 11 September 2026, retrained it on historical scans, handwriting, receipts and hard tables, then added a reinforcement-learning stage aimed at hallucinated text and runaway repetition.
The design is unusual. Most OCR models pair a vision encoder with a separate text decoder; Falcon OCR feeds image patches and text tokens into one 22-layer Transformer from the first layer, with image tokens attending in both directions and text decoded left to right. The Hugging Face API counts 270 million parameters. An optional mode runs Baidu's PP-DocLayoutV3 detector first, adding about 30 million more.
Is Falcon OCR worth it?
Falcon OCR's case is size against accuracy, and on TII's numbers the trade is good. Version 1.5 scores 82.4 on olmOCR-bench end to end, against 78.7 for PaddleOCR-VL-1.6 and 73.3 for DeepSeek-OCR 2, and only 1.4 points behind Chandra OCR 2 at 83.8. The comparison table is TII's, the card does not say whether rival scores were rerun or copied, and the average leaves out olmOCR's header and footer category, so it cannot be set beside officially reported olmOCR averages such as Marker's 76.0.
Tables are the weak spot. On OmniDocBench end-to-end mode reaches 86.47 overall, but table structure scores 69.51 TEDS against 90.24 for PaddleOCR-VL-1.6 in the same table. Switching on the layout pipeline lifts the overall score to 93.15 and tables to 88.53, at the cost of a second model, while its olmOCR average falls to 80.6. The card recommends end-to-end mode for most pages and the pipeline only for dense newspapers, which leaves table-heavy documents in an awkward middle.
The operating numbers are the most concrete of any model in this group. On one A100-80GB with vLLM, the full layout-and-OCR pipeline processes 2.9 images a second at 5,825 tokens a second, and the weights are a single 1.08 GB file stored in float32. TII ships a Docker image with an OpenAI-compatible endpoint and a PDF upload route, and documents running both services on a single GPU.
The costs are the usual ones for a custom architecture. Loading needs trust_remote_code and PyTorch 2.5 or later for FlexAttention, and the first call is slow while torch.compile builds kernels. The card lists degraded scans and very small text as open problems; old scans score 41.8 on olmOCR, ahead of most rivals but still the weakest category.
Where it earns its place
A platform engineer at an expense-management start-up
Read receipts and photographed invoices at high volume on a small GPU budget
Run the vLLM Docker image with skip_layout set to true and send images to the upload route. The 270M model leaves room on one card for high concurrency; the card's A100 figure is 2.9 images a second with layout switched on.
An archivist digitising handwritten letters and historical registers
Transcribe mixed handwriting and faded print without a cloud service
Version 1.5 retrained on handwriting and historical scans, and the weights run on a workstation GPU. Spot-check the faintest pages, which the card names as a remaining weakness.
A data engineer parsing newspaper back issues for search
Keep reading order on dense multi-column pages
Call generate_with_layout so PP-DocLayoutV3 finds regions before Falcon OCR reads each crop. Give the layout model its own GPU for best throughput, as the Docker guide suggests.
Before you commit
- Setup
- pip install torch>=2.5, transformers, pillow and einops, then load tiiuae/Falcon-OCR with trust_remote_code; or docker run ghcr.io/tiiuae/falcon-ocr:latest, which exposes vLLM on port 8000 and the parsing pipeline on 5002.
- Integrations
- Python generate() with 11 output categories, an optional layout pipeline, an OpenAI-compatible vLLM endpoint and a multipart upload route that accepts images and multi-page PDFs.
- Data and privacy
- Self-hosted, so documents stay on your hardware. Apache 2.0 permits commercial use; loading runs Python from the model repository.
- Cost at scale
- Weights are free; TII reports 2.9 images and 5,825 tokens a second on one A100-80GB through the full pipeline, so one card covers roughly 250,000 images a day at that rate. No hosted price exists.
- Limits
- Weak table structure end to end; maximum model length 16,384 tokens; degraded scans and very small glyphs flagged by the card; the comparison table is TII's own.
What’s good
- 270M parameters, a single 1.08 GB weights file
- 82.4 on olmOCR-bench end to end, per TII
- 2.9 images a second on one A100 with layout
- Apache 2.0, with an official vLLM Docker image
What’s not
- Comparison table is TII's own, and the olmOCR average omits headers and footers
- 69.51 table TEDS in end-to-end mode
- Needs trust_remote_code and PyTorch 2.5+
- Degraded scans and tiny text remain weak
Best for
Teams running high page volumes of receipts, forms, scans and handwriting on modest GPUs, where text matters more than table structure.
Falcon OCR alternatives
Falcon OCR: common questions
- Is Falcon OCR worth it?
- Falcon OCR reads text nearly as well as models ten times its size for the cost of a 270M-parameter pass, provided you can live without strong table structure. We rate it 4.0 out of 5.
- How much does Falcon OCR cost?
- Free under Apache 2.0. No hosted API; TII publishes a self-hosted vLLM Docker image.
- What are the downsides of Falcon OCR?
- Comparison table is TII's own, and the olmOCR average omits headers and footers; 69.51 table TEDS in end-to-end mode; needs trust_remote_code and PyTorch 2.5+; degraded scans and tiny text remain weak.
- Who is Falcon OCR best for?
- Teams running high page volumes of receipts, forms, scans and handwriting on modest GPUs, where text matters more than table structure.
- How much does Falcon OCR cost for a team or in production?
- Weights are free; TII reports 2.9 images and 5,825 tokens a second on one A100-80GB through the full pipeline, so one card covers roughly 250,000 images a day at that rate. No hosted price exists.
- Is Falcon OCR safe for company data?
- Self-hosted, so documents stay on your hardware. Apache 2.0 permits commercial use; loading runs Python from the model repository.
- What are the best Falcon OCR alternatives?
- The closest alternatives we have reviewed are PaddleOCR-VL-1.6, Surya OCR 2.5, DeepSeek-OCR 2, FireRed-OCR and dots.ocr.
Answers checked Nov 2, 2026.
Make Falcon OCR? Show this rating on your site
