Hayz Labs · https://hayzlabs.com/reviews/top/document-tools/
reviews · top 20 list · checked 18 Oct
Top tools for turning documents into text for AI
Tools that turn PDFs, scans, Office files and web pages into clean text or Markdown, so an AI model can read them.
Best: MinerU (80, free). 7 of 19 scored.
Scored ones first, highest score first; on a tie, the better result in the independent test. Unscored ones follow, by 30-day PyPI downloads, then GitHub stars. Open the Table to see the number that placed each one and how it was scored.
- 1MinerUOpenDataLab (Shanghai AI Lab)
Most accurate on CPU in the one fresh outside test; its own licence adds terms for very large services
Best for Scientific papers and complex layouts, parsed on an ordinary computer processorYou give up Online services must credit it; services above 100M monthly users need a paid licence
80ChecklistFree - 2DoclingDocling project (IBM origin)
Popular MIT-licensed parser from IBM Research with very frequent releases; the fresh outside test covered PDFs only
Best for Turning PDFs, Office files and web pages into clean text with good tablesYou give up Placed third of 7 on hard PDFs in the outside test, behind MinerU and Marker
70ChecklistFree - 3LiteParseLlamaIndex
Very fast, tiny local parser for simple PDFs; well behind the leaders on accuracy
Best for Simple PDFs where speed and low memory matter more than accuracyYou give up Scored 40.6 to MinerU's 74.6 in the outside test, so hard pages may come out wrong
70ChecklistFree - 4UnstructuredUnstructured
Splits many file types into clean pieces for search pipelines; sends usage pings unless you opt out
Best for Search pipelines that take in many file types, not only PDFsYou give up Placed fifth of 7 on hard PDFs; the hosted platform is sold separately
70ChecklistFree - 5MarkItDownMicrosoft
Quick way to turn Office files and web pages into Markdown; scored last on hard PDFs, so use another for scans
Best for Quick Markdown copies of Word, Excel and web pages to feed a modelYou give up Its Office and web converters were not part of the outside test; simpler than Docling
70ChecklistFree - 6MarkerDatalab
Second most accurate on CPU in the outside test but slow there; the weights need a licence above $5M funding or revenue
Best for Small firms wanting accurate PDF conversion, with optional AI help on hard pagesYou give up Slow on an ordinary processor: 9% of pages timed out; no release since 20 July 2026
60ChecklistFree - 7PyMuPDF4LLMArtifex
Fast, light PDF text extraction; AGPL licence, so closed-source business use needs a paid licence
Best for Open-source projects that need quick PDF text with no usage data sentYou give up Placed sixth of 7 on hard PDFs; no release since 6 August 2026
60ChecklistFree - 8Azure Document IntelligenceMicrosoft
Microsoft's managed OCR and form extraction; a natural pick if your company already runs on Azure
Best for Pulling fields from invoices, receipts and IDs inside Microsoft's cloudYou give up Layout and prebuilt reading cost $10 per 1,000 pages after 500 free; not scored yet
Not scored yetFree - 9ReductoReducto
Paid parsing and extraction API priced per page, pay as you go with no monthly plan
Best for Messy PDFs, tables and forms in finance work, with no monthly plan to commit toYou give up Parse costs $10 per 1,000 pages to Mistral OCR's $4; free allowance stated unclearly
Not scored yetFree - 10Unstructured PlatformUnstructured
Hosted version of Unstructured with a 10,000-page free start, for teams that don't want to run it
Best for Teams already using Unstructured who would rather not host itYou give up $15 per 1,000 pages after the first 10,000, while the open library is free to self-host
Not scored yetFree - 11LlamaParseLlamaIndex
LlamaIndex's paid parsing service with a free monthly allowance; pricier modes use AI models on hard pages
Best for LlamaIndex users who want hosted parsing with a cheap mode and a careful modeYou give up Cost per page depends on the mode you pick; not scored yet: no recent independent test
Not scored yetFree - 12Google Document AIGoogle
Google Cloud's managed OCR and form parsing; low-cost bulk OCR if you already use Google Cloud
Best for Low-cost bulk text reading from scans for teams on Google CloudYou give up Form parsing costs $30 per 1,000 pages; not scored yet: no recent independent test
Not scored yetFree - 13PaddleOCR (with PaddleOCR-VL)Baidu
Baidu's free OCR toolkit, including the small PaddleOCR-VL model for whole-page parsing
Best for Reading text from scans for free, with a small model for whole pagesYou give up Scores we found are the maker's own or compiled; not scored yet
Not scored yetFree - 14Amazon TextractAmazon
Amazon's managed OCR with table and form extraction; fits teams already on AWS
Best for Teams on AWS who mainly need plain text, its cheapest mode at $1.50 per 1,000 pagesYou give up Forms cost $50 per 1,000 pages, the highest rate on this list; not scored yet
Not scored yetPrice varies - 15OpenDataLoader PDFOpenDataLoader project
Free Apache-2.0 PDF to Markdown or JSON parser that rose fast in 2026; little outside testing yet
Best for Developers wanting a free, permissively licensed PDF parser to tryYou give up Benchmark claims on its page are its own; not scored yet: no independent test
Not scored yetFree - 16Datalab APIDatalab
Paid API from the makers of Marker and Chandra, for teams that don't want to run their own GPUs
Best for Teams that want Marker-style PDF conversion as a pay-per-page serviceYou give up Accurate mode costs $10 per 1,000 pages; the Team plan is $400 a month; not scored yet
Not scored yetFree - 17olmOCRAi2
Ai2's free toolkit for turning large PDF collections into clean text; needs a GPU
Best for Bulk conversion of large PDF archives on your own graphics cardYou give up Needs a graphics card; its only scores come from Ai2, which also runs that benchmark
Not scored yetFree - 18ChandraDatalab
Datalab's open OCR model for scans; free weights only under $2M revenue and not for rival APIs
Best for Small companies reading scanned pages with a free open modelYou give up Weights need a licence above $2M revenue and can't power a rival service; own scores only
Not scored yetFree - 19Mistral OCRMistral AI
Low-cost OCR API from Mistral; the current model is OCR 4.1
Best for Low-cost bulk scan reading, with batch jobs at $2 per 1,000 pagesYou give up Top score claim is Mistral's own; hosted only, with self-hosting just for enterprise
Not scored yetPay as you go
| # | Name | Score | What placed it | Price | Best for | You give up |
|---|---|---|---|---|---|---|
| 1 | MinerUOpenDataLab (Shanghai AI Lab) · Open source | 80Our checklist | About 112,000 downloads a month from PyPI, Python's package storeWhat it rests onTheir figure: PyPI mineru 112k downloads in 30 days to 2026-10-01; olmOCR-Bench on a 4-core CPU 74.6 (#1 of 7, pdf-cpu-bench). checklist: job 2 (pdf-cpu-bench run 2026-10-07 tested 4.0.10, newest at the time; 74.6, #1 of 7 on CPU), model 0 (no outside rating), value 2 (free), privacy 2 (opt-out telemetry, no content), limits 2 = 8 Report a mistake | Free (custom MinerU licence) | Scientific papers and complex layouts, parsed on an ordinary computer processor | Online services must credit it; services above 100M monthly users need a paid licence |
| 2 | DoclingDocling project (IBM origin) · Open source | 70Our checklist | About 2.9 million downloads a month from PyPI, Python's package storeWhat it rests onTheir figure: PyPI docling 2.91M downloads in 30 days to 2026-10-01; olmOCR-Bench on a 4-core CPU 53.4 (#3 of 7). checklist: job 1 (pdf-cpu-bench run 2026-10-07 tested 2.134.0, newest at the time; 53.4, #3 of 7), model 0, value 2, privacy 2, limits 2 = 7 Report a mistake | Free (open source) | Turning PDFs, Office files and web pages into clean text with good tables | Placed third of 7 on hard PDFs in the outside test, behind MinerU and Marker |
| 3 | LiteParseLlamaIndex · Open source | 70Our checklist | About 372,000 downloads a month from PyPI, Python's package storeWhat it rests onTheir figure: PyPI liteparse 372k downloads in 30 days to 2026-10-01; olmOCR-Bench on a 4-core CPU 40.6 (#4 of 7). checklist: job 1 (pdf-cpu-bench run 2026-10-07 tested 2.15.1, newest at the time; 40.6, 0.61 s/page, 0.36 GB RAM), model 0, value 2, privacy 2, limits 2 = 7 Report a mistake | Free (open source) | Simple PDFs where speed and low memory matter more than accuracy | Scored 40.6 to MinerU's 74.6 in the outside test, so hard pages may come out wrong |
| 4 | UnstructuredUnstructured · Open source | 70Our checklist | About 2.6 million downloads a month from PyPI, Python's package storeWhat it rests onTheir figure: PyPI unstructured 2.58M downloads in 30 days to 2026-10-01; olmOCR-Bench on a 4-core CPU 38.0 (#5 of 7). checklist: job 1 (pdf-cpu-bench run 2026-10-07 tested 0.27.16, newest at the time; 38.0 hi_res), model 0, value 2, privacy 2 (usage pings, opt-out, no content), limits 2 = 7 (open library only) Report a mistake | Free (open source) | Search pipelines that take in many file types, not only PDFs | Placed fifth of 7 on hard PDFs; the hosted platform is sold separately |
| 5 | MarkItDownMicrosoft · Open source | 70Our checklist | About 16 million downloads a month from PyPI, Python's package storeWhat it rests onTheir figure: PyPI markitdown 15.99M downloads in 30 days to 2026-10-01; olmOCR-Bench on a 4-core CPU 29.3 (#7 of 7). checklist: job 1 (pdf-cpu-bench run 2026-10-07 tested the current 0.1.8: 29.3, lowest of 7 on PDFs, but fastest and lightest; its Office and HTML converters are not covered by that test), model 0 (no outside rating), value 2, privacy 2, limits 2 (0.1.8 on 2026-09-21; recheck commits after 2026-10-21) = 7 Report a mistake | Free (open source) | Quick Markdown copies of Word, Excel and web pages to feed a model | Its Office and web converters were not part of the outside test; simpler than Docling |
| 6 | MarkerDatalab · Open source | 60Our checklist | About 121,000 downloads a month from PyPI, Python's package storeWhat it rests onTheir figure: PyPI marker-pdf 121k downloads in 30 days to 2026-10-01; olmOCR-Bench on a 4-core CPU 58.3 (#2 of 7). checklist: job 1 (pdf-cpu-bench run 2026-10-07 tested the current 2.0.0: 58.3, #2 of 7, but 9% of pages hit the 5-minute timeout on CPU), model 0 (no outside rating), value 2 (free), privacy 2, limits 1 (no release since 2026-07-20 and recent commits could not be confirmed; lower score taken) = 6 Report a mistake | Free (code Apache-2.0); weights licence needed above $5M funding or revenue | Small firms wanting accurate PDF conversion, with optional AI help on hard pages | Slow on an ordinary processor: 9% of pages timed out; no release since 20 July 2026 |
| 7 | PyMuPDF4LLMArtifex · Open source | 60Our checklist | About 9.2 million downloads a month from PyPI, Python's package storeWhat it rests onTheir figure: PyPI pymupdf4llm 9.23M downloads in 30 days to 2026-10-01; olmOCR-Bench on a 4-core CPU 37.5 (#6 of 7). checklist: job 1 (pdf-cpu-bench run 2026-10-07 tested the current 1.28.2: 37.5, 0.85 s/page), model 0, value 2 (free; AGPL), privacy 2 (no telemetry), limits 1 (no release since 2026-08-06 and recent commits could not be confirmed; lower score taken) = 6 Report a mistake | Free (AGPL); paid commercial licence from Artifex | Open-source projects that need quick PDF text with no usage data sent | Placed sixth of 7 on hard PDFs; no release since 6 August 2026 |
| 8 | Azure Document IntelligenceMicrosoft · Closed | Not scored | About 11 million downloads a month from PyPI, Python's package storeNotesTheir figure: PyPI azure-ai-documentintelligence 11.30M downloads in 30 days to 2026-10-01. not scored: no independent test of the current service in the last 90 days Report a mistake | 500 pages/mo free; Read $1.50, prebuilt and layout $10 per 1,000 pages | Pulling fields from invoices, receipts and IDs inside Microsoft's cloud | Layout and prebuilt reading cost $10 per 1,000 pages after 500 free; not scored yet |
| 9 | ReductoReducto · Closed | Not scored | About 4.4 million downloads a month from PyPI, Python's package storeNotesTheir figure: PyPI reductoai 4.38M downloads in 30 days to 2026-10-01. not scored: no independent test in the last 90 days; openbenchmarks.com contract test (2026-09-20) set aside: narrow task, publisher sells benchmarking to vendors Report a mistake | Parse $10, Extract $20 per 1,000 pages; free starter credit | Messy PDFs, tables and forms in finance work, with no monthly plan to commit to | Parse costs $10 per 1,000 pages to Mistral OCR's $4; free allowance stated unclearly |
| 10 | Unstructured PlatformUnstructured · Closed | Not scored | About 3.3 million downloads a month from PyPI, Python's package storeNotesTheir figure: PyPI unstructured-client 3.29M downloads in 30 days to 2026-10-01. not scored: no independent test of the hosted service in the last 90 days (the open library is scored separately) Report a mistake | First 10,000 pages free; then $15 per 1,000 pages | Teams already using Unstructured who would rather not host it | $15 per 1,000 pages after the first 10,000, while the open library is free to self-host |
| 11 | LlamaParseLlamaIndex · Closed | Not scored | About 3.3 million downloads a month from PyPI, Python's package storeNotesTheir figure: PyPI llama-cloud 3.25M downloads in 30 days to 2026-10-01 (SDK for all LlamaCloud services). not scored: no independent test in the last 90 days; openbenchmarks.com contract test (2026-09-20) set aside: narrow task, publisher sells benchmarking to vendors Report a mistake | Free 10K credits/mo; Starter $50/mo; Pro $500/mo; 1,000 credits = $1.25 | LlamaIndex users who want hosted parsing with a cheap mode and a careful mode | Cost per page depends on the mode you pick; not scored yet: no recent independent test |
| 12 | Google Document AIGoogle · Closed | Not scored | About 3.2 million downloads a month from PyPI, Python's package storeNotesTheir figure: PyPI google-cloud-documentai 3.24M downloads in 30 days to 2026-10-01. not scored: no independent test of the current service in the last 90 days Report a mistake | OCR $1.50 per 1,000 pages (first 1,000 free); Layout Parser $10 per 1,000 | Low-cost bulk text reading from scans for teams on Google Cloud | Form parsing costs $30 per 1,000 pages; not scored yet: no recent independent test |
| 13 | PaddleOCR (with PaddleOCR-VL)Baidu · Open source | Not scored | About 1.3 million downloads a month from PyPI, Python's package storeNotesTheir figure: PyPI paddleocr 1.29M downloads in 30 days to 2026-10-01. not scored: no independent test in the last 90 days; OmniDocBench and olmOCR-Bench figures seen are maker-reported or compiled Report a mistake | Free (open source) | Reading text from scans for free, with a small model for whole pages | Scores we found are the maker's own or compiled; not scored yet |
| 14 | Amazon TextractAmazon · Closed | Not scored | About 906,000 downloads a month from PyPI, Python's package storeNotesTheir figure: PyPI amazon-textract-textractor 906k downloads in 30 days to 2026-10-01 (most calls go through boto3). not scored: no independent test of the current service in the last 90 days Report a mistake | Text $1.50, Tables $15, Forms $50 per 1,000 pages (US West) | Teams on AWS who mainly need plain text, its cheapest mode at $1.50 per 1,000 pages | Forms cost $50 per 1,000 pages, the highest rate on this list; not scored yet |
| 15 | OpenDataLoader PDFOpenDataLoader project · Open source | Not scored | About 155,000 downloads a month from PyPI, Python's package storeNotesTheir figure: PyPI opendataloader-pdf 155k downloads in 30 days to 2026-10-01; GitHub 29.7k stars. not scored: no independent test in the last 90 days; benchmark claims on its page are its own Report a mistake | Free (open source) | Developers wanting a free, permissively licensed PDF parser to try | Benchmark claims on its page are its own; not scored yet: no independent test |
| 16 | Datalab APIDatalab · Closed | Not scored | About 152,000 downloads a month from PyPI, Python's package storeNotesTheir figure: PyPI datalab-python-sdk 152k downloads in 30 days to 2026-10-01. not scored: no independent test in the last 90 days; openbenchmarks.com contract test (2026-09-20) set aside: narrow task, publisher sells benchmarking to vendors Report a mistake | Convert $4 per 1,000 pages (accurate $10); $10 to $20 free monthly; Team $400/mo | Teams that want Marker-style PDF conversion as a pay-per-page service | Accurate mode costs $10 per 1,000 pages; the Team plan is $400 a month; not scored yet |
| 17 | olmOCRAi2 · Open source | Not scored | About 19,200 GitHub stars: popular, which isn't the same as goodNotesTheir figure: GitHub 19.2k stars (PyPI olmocr under the top 15,000). not scored: olmOCR-Bench scores for it come from its own maker, which also runs that benchmark Report a mistake | Free (open source) | Bulk conversion of large PDF archives on your own graphics card | Needs a graphics card; its only scores come from Ai2, which also runs that benchmark |
| 18 | ChandraDatalab · Open source | Not scored | About 12,500 GitHub stars: popular, which isn't the same as goodNotesTheir figure: GitHub 12.5k stars (PyPI chandra-ocr under the top 15,000). not scored: no independent test in the last 90 days; its olmOCR-Bench score is self-reported Report a mistake | Free; weights licence needed above $2M revenue | Small companies reading scanned pages with a free open model | Weights need a licence above $2M revenue and can't power a rival service; own scores only |
| 19 | Mistral OCRMistral AI · Closed | Not scored | No OCR-only usage figure (the mistralai SDK covers all its APIs)NotesTheir figure: no OCR-only usage figure (the mistralai SDK covers all its APIs). not scored: no independent test in the last 90 days; the top OlmOCRBench claim is its own; openbenchmarks.com contract test (2026-09-20) set aside: narrow task, publisher sells benchmarking to vendors Report a mistake | API $4 per 1,000 pages ($5 with annotations) | Low-cost bulk scan reading, with batch jobs at $2 per 1,000 pages | Top score claim is Mistral's own; hosted only, with self-hosting just for enterprise |
Nothing matches all of that. , or try everything.
How we ranked this
Order. Scored ones first, highest score first; on a tie, the better result in the independent test. Unscored ones follow, by 30-day PyPI downloads, then GitHub stars.
How we score this list. Our checklist. The one fresh independent test (olmOCR-Bench on a 4-core CPU, run 7 Oct 2026) covers seven open tools, so paid services and GPU models are Not scored. The full rules are on How we score. Open "How it was scored" on any row to see the working.
What we did not do. We did not run any of these ourselves. Every number comes from the public source linked on its row, checked on 18 Oct 2026.
Spot a mistake? A wrong rank, price or link: tell us and we will check it against the source. Corrections are dated in News.
Reviews are drafted with help from AI models, including Anthropic's Claude. Model scores come only from outside leaderboards and independent evaluations, by a written rule.