Core capabilities
What deepseek ocr2 delivers
Six OCR pillars tuned for speed, accuracy, and production reliability.
Workflow
From scan to structured data in minutes
Ingest
Upload PDFs, images, or scans via API or batch sync.
Detect & parse
OCR2 reads text, tables, and layout in one pass.
Validate
Check confidence scores and apply domain rules.
Export
Ship JSON, CSV, or searchable PDF to your pipeline.
Performance
Accuracy without slowing you down
Fast inference with deterministic layout even under heavy loads.
97.4%
Table accuracy
Benchmarked on complex invoices and reports.
30+
Multi-language
Consistent OCR for mixed scripts.
12k pages/hr
Batch throughput
Measured on standard A4 scans.
Research notes
Paper-driven read on OCR2
From the paper, OCR2 centers on causal visual flow, token reordering, and document-first evaluation.
DeepEncoder V2 swaps CLIP for an LLM-style encoder and adds causal flow queries.
Visual tokens keep bidirectional attention; causal queries are causal, and only their outputs go to the decoder.
The paper reports a 3.73% overall gain on OmniDocBench v1.5 over DeepSeek-OCR.
Visual token budget is constrained between 256 and 1120 for cost vs fidelity.
Training uses ~80% OCR data with three stages: pretraining, query enhancement, decoder specialization.
Dynamic resolution inference supports mixed-size documents and scans.
TextEdit
Edit distance for text correctness.
Formula CDM
Formula consistency metric.
Table TEDS
Table structure similarity score.
R-order Edit
Reading order edit distance.
Prompt modes
Layout-preserving
<image>\n<|grounding|> Convert the document to markdown.
Plain OCR
<image>\nFree OCR.
Use cases
Where deepseek ocr2 works best
Built for finance, logistics, customer support, and knowledge teams.
Invoice automation
Extract totals, line items, and tax fields reliably.
Compliance archives
Search and audit long-tail documents with confidence.
Shipping & logistics
Normalize manifests, customs docs, and labels.
Knowledge digitization
Turn scanned manuals into searchable references.
FAQ
Common questions
Can deepseek ocr2 handle mixed languages?
Yes. OCR2 detects and normalizes mixed scripts automatically.
Do you support private deployment?
Private and hybrid options are available with audit controls.
How do we evaluate accuracy?
Start with real documents and compare field-level precision.
What file types are supported?
Images, PDFs, scans, and multi-page documents.
Ready to ship
Start with deepseek ocr2 today
Move from scans to structured data with confidence.