Skip to main content
Updated July 12, 2026
Document OCR and parsing models turn PDFs, scans, and images into clean, structured text software can use. The catch: a model can nail clean invoices and fall apart on dense tables or handwriting. The 15 picks below are ordered by normalized ParseBench score and practical access tradeoffs.

​
Best Document OCR & Parsing Models

​
LlamaParse

LlamaIndex
Agentic parsing for RAG
Visit LlamaIndex
The strongest all-round parser here, turning messy PDFs into clean, RAG-ready Markdown that holds structure where cheaper tools quietly drop it.
Score 100Price License ProprietaryParser type Specialized parser
  • Its agentic mode runs multi-step vision reasoning to rebuild tables, charts, and multi-column layouts into clean Markdown ready for a RAG pipeline.
  • On dense enterprise pages it holds structure where lighter parsers drop rows or scramble reading order, which makes it a dependable default.
  • Even the best parsers still omit or hallucinate content on a small share of pages, so high-stakes fields need a verification pass.
  • The top mode is pricey per page, and if you want field-level citations for audit, Reducto is built more directly for that.

​
KDL-Frontier-Parser-nano

KoreaDeep / KDLAI
Open-weight visual grounding
View on Hugging Face
The open-weight parser to beat when you need precise on-page coordinates, not just clean text, and you have a GPU to run it.
Score 88Price License Open weightParser type Open-weight VLM
  • It excels at visual grounding, locating exactly where each element sits on the page, which matters when you need to link extracted values back to their source region for review or highlighting.
  • As a compact open-weight model, it gives self-hosting teams frontier-level structure without sending documents to anyone else.
  • You need a high-end GPU and your own serving stack, so it is not a drop-in API.
  • General VLMs like Gemini 3 Flash are easier to call, and if you want self-hosting on lighter hardware, MinerU2.5-Pro is the easier route.
  • Run locally — If you have a high-end machine, you can run it with vLLM after downloading weights from Hugging Face.

​
Gemini 3 Flash

Google
Fast general-purpose parsing
Visit Google
The most capable general VLM for parsing at speed - not a purpose-built parser, but fast, cheap enough for volume, and rarely embarrassing.
Score 86Price License ProprietaryParser type VLM API
  • A strong all-purpose route when you want parsing plus reasoning in one call: ask questions, extract fields, and summarize in the same request.
  • Its very large context handles long documents in one pass, and you can dial visual detail up or down to trade accuracy against cost per page.
  • As a general model it trails purpose-built parsers on the hardest tables and dense layouts, where LlamaParse and Reducto pull ahead.
  • For steady structured extraction at volume, a dedicated OCR API like Mistral OCR 4 can be more predictable and cheaper.

​
Infinity-Parser2-Pro

infly
Highest-accuracy open weights
View on Hugging Face
The open-weight pick when raw parsing accuracy matters most, especially across English and Chinese documents, if you can host it yourself.
Score 85Price License Open weightParser type Open-weight VLM
  • Reinforcement-tuned specifically for parsing, it is one of the most accurate open-weight models for tables, formulas, and reading order, and it handles English and Chinese documents equally well.
  • Self-hosting keeps sensitive files in-house, and a lighter Flash variant trades some accuracy for faster throughput when you need it.
  • It wants a high-end GPU and hands-on serving, so it is not a fast start for small teams.
  • If you want self-hosting on modest hardware, MinerU2.5-Pro or PaddleOCR-VL run more easily; skip it entirely if you would rather not host a model at all.

​
Reducto

Reducto
Auditable enterprise extraction
Visit Reducto
Built for regulated, high-stakes extraction where every value needs a citation, and the safe choice when a wrong field has real consequences.
Score 83Price License ProprietaryParser type Specialized parser
  • It re-examines low-confidence regions and returns bounding boxes, per-field citations, and confidence scores, so a reviewer can trace every extracted value back to the page.
  • That auditability, plus on-prem deployment and strong compliance support, makes it a natural fit for finance, insurance, and healthcare workflows.
  • It is among the priciest options per page, so it is overkill for casual or low-stakes parsing.
  • For clean Markdown to feed a RAG pipeline, LlamaParse scores higher for less money, and general VLMs cost far less when you do not need citations.

​
MinerU2.5-Pro

OpenDataLab
Local technical-document parsing
View on Hugging Face
The best open-weight parser you can actually run on normal hardware, and a standout on dense academic and technical documents.
Score 83Price License Open weightParser type Open-weight VLM
  • A compact model that punches well above its size on scientific and technical PDFs, where formulas, nested tables, and multi-column layouts come through cleanly.
  • Because it runs on a typical machine through the MinerU toolkit, you get strong parsing offline, with no per-page fees and nothing leaving your device.
  • It is a self-hosted toolkit, not a managed API, so you own setup, updates, and scaling.
  • For hands-off parsing, LlamaParse or Mistral OCR 4 are simpler, and for the very hardest enterprise layouts the top hosted parsers still hold an edge.
  • Run locally — You can run it locally with MinerU after downloading weights from Hugging Face.

​
Claude Fable 5

Anthropic
Reasoning-heavy extraction
Visit Anthropic
Reach for it when parsing bleeds into judgment: reading a document, reasoning over it, and extracting structured answers in one step.
Score 80Price License ProprietaryParser type VLM API
  • Its strength is document understanding, not just transcription. It follows complex instructions, reasons across pages, and returns structured output that reflects what the document means, not only what it says.
  • For messy, ambiguous documents that need interpretation rather than literal extraction, it is unusually reliable.
  • It is one of the most expensive options here and is a general model, not a dedicated parser, so for high-volume plain OCR it is hard to justify.
  • For pure layout and table extraction, LlamaParse and Mistral OCR 4 do more per dollar.

​
Chandra OCR 2

Datalab
Tables, forms, and handwriting
View on Hugging Face
A top open-weight OCR model for the ugly stuff - complex tables, dense forms, and handwriting - with a hosted option if you skip self-hosting.
Score 79Price License Open weightParser type Open-weight VLM
  • It handles the documents that break simpler OCR: intricate tables, structured forms, and handwriting, all while preserving full page layout.
  • Rare among open-weight models, it stays competitive with proprietary parsers, which makes it a strong choice when you want frontier-level extraction without a closed API.
  • Running the weights yourself needs a high-end GPU, so the hosted route is realistic for most teams.
  • On clean printed text it is close to lighter models like Surya OCR 2 that run on far less hardware, so save it for genuinely hard pages.

​
Datalab Parser

Datalab
Low-cost hosted parsing
Visit Datalab
A low-cost hosted parser from the team behind Surya and Chandra that quietly does the job and offers strong value for everyday document work.
Score 79Price License ProprietaryParser type Specialized parser
  • It delivers solid, well-structured parsing at one of the lowest hosted prices here, which makes it easy to run at volume without watching the meter.
  • For standard business documents like invoices, reports, and contracts, it hits a practical accuracy-to-cost balance most projects can build on.
  • It is a pragmatic middle option, not a top scorer, so the hardest layouts still favor LlamaParse or Reducto.
  • And because it is a managed API, it does not give you the offline control of the open-weight models from the same team.

​
Mistral OCR 4

Mistral AI
High-volume OCR at scale
Visit Mistral
A fast, low-cost OCR API built for volume, and the pick when you need to process a lot of pages cheaply and reliably.
Score 76Price License ProprietaryParser type Cloud OCR API
  • It collapses OCR, layout, and structured extraction into a single fast call, with bounding boxes and structured output that drop cleanly into a pipeline.
  • Low per-page cost and steady throughput make it well suited to high-volume workloads where you need predictable results without managing infrastructure.
  • It stumbles on math, scientific notation, and complex multi-column pages, and outputs sometimes need manual review.
  • For those harder documents, LlamaParse or MinerU2.5-Pro are safer, and general VLMs handle unusual layouts more gracefully.

​
GPT-5.5

OpenAI
Versatile document reasoning
Visit OpenAI
The versatile generalist - not the sharpest on pure OCR benchmarks, but flexible, strong on handwriting, and easy to fold into wider workflows.
Score 76Price License ProprietaryParser type VLM API
  • As a frontier general model it parses, reasons, and answers questions about a document in one call, and it is among the better options for handwriting.
  • When parsing is one step inside a larger reasoning or agent task, doing it all in a single model keeps the pipeline simple.
  • On raw parsing accuracy it sits mid-pack, behind dedicated parsers like LlamaParse and even strong open-weight models.
  • It is also expensive for high-volume OCR, so for pure extraction at scale, Mistral OCR 4 or a self-hosted parser makes more sense.

​
PaddleOCR-VL

Baidu / PaddlePaddle
Multilingual local parsing
View on Hugging Face
An ultra-compact open-weight parser with unusually broad language coverage that runs on ordinary hardware, making it a strong pick for multilingual work.
Score 75Price License Open weightParser type Open-weight VLM
  • Despite its tiny size, it delivers strong document parsing across a very wide set of languages, which makes it a standout for non-English and mixed-language documents.
  • It runs on a typical machine, so you get multilingual extraction offline, with no per-page cost and full control over your data.
  • The compact size shows on the most complex enterprise layouts, where larger parsers pull ahead.
  • If you need the highest ceiling and can host bigger weights, Infinity-Parser2-Pro or MinerU2.5-Pro are stronger; for hands-off use, a hosted API is simpler.

​
Surya OCR 2

Datalab
Lightweight local OCR
View on Hugging Face
The lightweight local workhorse, small enough to run almost anywhere including CPU and Apple Silicon, while still covering dozens of languages.
Score 71Price License Open weightParser type Open-weight VLM
  • It rolls layout analysis, OCR, and table recognition into one small model that runs on modest hardware, even without a dedicated GPU.
  • With coverage across dozens of languages and a genuinely lightweight footprint, it is one of the easiest ways to get solid offline OCR onto a normal laptop.
  • Its small size caps accuracy on complex tables and dense layouts, where Chandra OCR 2 or MinerU2.5-Pro do better.
  • The weights also carry usage terms worth checking before you ship it in a commercial product.

​
Azure Document Intelligence

Microsoft
Prebuilt form extraction
Visit Microsoft
A mature cloud OCR service with strong prebuilt models for forms and invoices, dependable for structured fields but less so for open-ended parsing.
Score 64Price License ProprietaryParser type Cloud OCR API
  • Years of refinement show in its prebuilt extractors for invoices, receipts, and IDs, plus reliable handling of printed text, forms, and tables.
  • For teams that need structured fields out of standardized business documents with minimal custom work, it is a proven, well-supported option.
  • It is built for structured field extraction, not the semantic, RAG-ready parsing that newer VLM parsers do best, so it trails them on complex or free-form layouts.
  • For clean Markdown from messy documents, LlamaParse or Gemini 3 Flash are stronger.

​
AWS Textract

Amazon Web Services
Forms and table extraction
Visit AWS
A dependable older-generation OCR service for clean forms and tables, now clearly outclassed on anything requiring semantic document understanding.
Score 47Price License ProprietaryParser type Cloud OCR API
  • For structured, well-scanned documents such as forms with key-value pairs and bordered tables, it is stable, scalable, and predictable.
  • If your inputs are clean and your needs are literal extraction rather than layout reconstruction, it does that job reliably at production scale.
  • It sits at the bottom on semantic parsing. It reads text but does not reconstruct document structure or meaning the way modern VLM parsers do.
  • For complex layouts, RAG-ready output, or messy scans, nearly everything above it does more.

​
How to Choose

When choosing between these models, consider:
  • Access: First decide whether you want a hosted app, an API call, or a model you run yourself, because that choice drives cost, privacy, latency, and how much setup you own. Proprietary parsers are the fastest to start; open-weight models keep documents on your own hardware.
  • Quality: We use the ParseBench overall score as the main measure. It tests how well parsed output preserves tables, charts, content faithfulness, semantic formatting, and on-page visual grounding across real enterprise documents, not just whether the text looks similar to a reference.
  • Price: We compare on USD per 1,000 pages processed, the cleanest way to line up hosted parsers. Self-hosted open-weight models carry no per-page fee, but you pay in hardware and setup instead.
  • Parser Type: The field splits into specialized parser APIs, general VLM APIs, open-weight VLMs, and cloud OCR APIs. Specialized parsers and open-weight VLMs lead on hard layouts, cloud OCR APIs stay steady on clean structured forms, and general VLMs add reasoning but are not purpose-built.

​
Other Models We Considered

Gemini 3.5 Flash (Google) — Fast and capable, but overlaps closely with Gemini 3 Flash.Gemini 3.1 Pro (Google) — Stronger for reasoning-heavy parsing, but pricier and slower than Gemini 3 Flash.Extend (Extend) — Capable extraction API, but narrower than the leading parsers here.Nanonets OCR-3 (Nanonets) — Popular hosted OCR for extraction, but outscored by the main picks.Qwen3-VL-8B-Instruct (Alibaba) — A solid open multimodal baseline, but not a purpose-built parser.Google Document AI (Google) — A familiar cloud baseline for structured docs, now behind newer parsers.Dots.mocr (RedNote HiLab) — An open OCR model for experiments, but well behind the leaders.Docling Models (IBM) — A handy offline conversion toolkit, but weaker on complex layouts.LandingAI ADE (LandingAI) — A hosted extraction platform, but low parsing accuracy on hard documents.DeepSeek-OCR-2 (DeepSeek) — Interesting for document compression, but low general parsing accuracy.

​
Frequently Asked Questions

LlamaParse in its agentic mode is the strongest all-round pick. It reconstructs complex layouts into clean, RAG-ready Markdown more reliably than anything else, and it is available as a simple API. If you need every extracted value to carry a citation for audit, Reducto is the more specialized choice.
For a hosted default, LlamaParse is the safest starting point. If you are processing large volumes and want to keep costs down, Mistral OCR 4 or Datalab Parser give you most of the quality for far less per page. Test two or three on your own documents before committing.
MinerU2.5-Pro and PaddleOCR-VL are the standouts because they run well on a typical machine, MinerU for technical documents and PaddleOCR-VL for multilingual work. Surya OCR 2 is the lightest option, even on CPU or Apple Silicon. KDL-Frontier-Parser-nano, Infinity-Parser2-Pro, and Chandra OCR 2 score higher but need a high-end GPU.
Use a dedicated parser when parsing is the whole job, since purpose-built models handle hard tables and dense layouts more reliably. Reach for a general VLM when parsing is one step inside a larger reasoning or agent task and you want everything in a single call.
Roughly, but not perfectly. Scores predict which models handle complex layouts, tables, and faithfulness well, yet performance swings with your specific document types, scan quality, and languages. Even the best parsers miss or invent content on a small share of pages, so verify critical fields and run a short test on your own files.
They are still fine for clean, structured forms and key-value extraction. But if you need semantic, RAG-ready output from messy or complex documents, a VLM parser like LlamaParse, Gemini 3 Flash, or an open-weight model like MinerU2.5-Pro will serve you much better.
Three things: how you want to access it (app, API, or self-hosted), the kind of documents you actually process, and your tolerance for cost versus accuracy. Match the model to your hardest real documents, not the cleanest ones, because that is where the differences show up.