olmOCR-2

olmOCR-2 is the Allen Institute for AI’s open-weight document OCR model, released Apache 2.0 in October 2025. A 7B Qwen2.5-VL fine-tune trained specifically on document-anchored data — academic papers, legal documents, historical scans, and handwriting — with GRPO reinforcement on unit-test rewards that target equation conversion, table-cell accuracy, and reading order.

We host it on our GPUs as the premium open-weight alternative to GLM-OCR for documents leaning Western / academic / legal — where olmOCR-2’s training data and benchmark profile give it the edge.

When to pick olmOCR-2 over GLM-OCR

• Academic papers — multi-column, equations, citations
• Legal documents and historical scans — reading order matters
• Handwritten content — heavily represented in olmOCR-2’s training set
• Sovereign procurement — Ai2 is a Seattle non-profit; if your buyer requires a non-PRC origin open-weight model, this is it
• You’ll tolerate a 1.67× per-page upcharge for the accuracy gain

When to pick GLM-OCR instead

• High-volume Chinese + English enterprise paperwork — invoices, receipts, mixed business documents. GLM-OCR is purpose-built for that and ~40% cheaper, with a volume tier above 500k pages/month.

When to pick dots.ocr instead

• Nepali, Hindi, Bengali, or other Devanagari / Indic script documents — dots.ocr is the multilingual specialist in the catalog.

Pricing

EUR 0.0025 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 (LaTeX equations, HTML tables), JSON with caller-supplied schema, or interleaved text + bounding boxes for audit pipelines.

Benchmark honesty

olmOCR-2 sits around 85.7 on OmniDocBench end-to-end and 82.4 on olmOCR-Bench — comfortably ahead of Mistral OCR (78), Marker, and most generic VLMs. It is not the highest-scoring model on OmniDocBench overall (MinerU 2.5 Pro and GLM-OCR both score higher there). We picked olmOCR-2 specifically for its license (pure Apache 2.0, no MAU caps, no attribution riders, no anti-compete clauses) and provenance (Ai2 non-profit, US origin), not for the absolute top of the leaderboard.