Mistral AI
A strong hosted PDF-to-markdown API for mixed and scanned documents, with solid OCR, table recovery, and asset export but uneven structural fidelity.
Strong conversion engine, mixed structure preservation
Mistral AI handled all three tested PDFs through a fully automated API workflow and returned useful markdown in both consolidated and page-level formats. It did especially well on readable text, many financial tables, scanned OCR, and keeping charts or signature assets attached to page output. The main weakness is fidelity to original structure: heading hierarchy, TOC nesting, and the semantics of the hardest multi-level tables were not preserved consistently enough to treat the markdown as a perfect reconstruction.
In-Depth Review
Our detailed analysis of Mistral AI — features, performance, and real-world testing.
Feature-by-Feature Breakdown
Structured markdown exportReliable export packaging across all tested PDFs.▾
Feature tested: Structured markdown export
Result: Passed
Verdict: Reliable export packaging across all tested PDFs.
Expected behavior: Mistral AI returned markdown as downloadable output rather than a UI-only preview. In the hybrid earnings report, table-heavy financial report, and scanned research paper tests, the export pattern included a full-document markdown file plus page-level outputs, which makes it easier to inspect one page at a time or ingest the whole document at once.
Test case: PDF document → File
Input type: PDF document
Input used: Input artifact (PDF document): 84-page hybrid earnings report with native text, financial tables, charts, and a scanned signature page. — llamaparse-hybrid-earnings-pdf-1.pdf
Observed output: Output artifact (File): Mistral AI returned the hybrid report as a downloadable ZIP containing markdown outputs rather than a single flat text dump. — mistral-ai-mistral-ai-hybrid-earnings-pdf-output-zip-3.zip
Input artifact: Input artifact (PDF document): 84-page hybrid earnings report with native text, financial tables, charts, and a scanned signature page. — llamaparse-hybrid-earnings-pdf-1.pdf
Output artifact: Output artifact (File): Mistral AI returned the hybrid report as a downloadable ZIP containing markdown outputs rather than a single flat text dump. — mistral-ai-mistral-ai-hybrid-earnings-pdf-output-zip-3.zip
What changed: PDF document transformed into File
Test case: PDF document → Image
Input type: PDF document
Input used: Input artifact (PDF document): Hybrid earnings report page export structure. — llamaparse-hybrid-earnings-pdf-1.pdf
Observed output: Output artifact (Image): Inside a page-level folder, the export included a markdown file plus separate image assets, showing that Mistral packages parsed content page by page instead of — mistral-ai-windows-explorer-page-folder.png
Input artifact: Input artifact (PDF document): Hybrid earnings report page export structure. — llamaparse-hybrid-earnings-pdf-1.pdf
Output artifact: Output artifact (Image): Inside a page-level folder, the export included a markdown file plus separate image assets, showing that Mistral packages parsed content page by page instead of — mistral-ai-windows-explorer-page-folder.png
What changed: PDF document transformed into Image
Test case: PDF document → Image
Input type: PDF document
Input used: Input artifact (PDF document): 18-page table-heavy financial report. — llamaparse-sumitomo-financial-pdf-1.pdf
Observed output: Output artifact (Image): The financial-report export contained a top-level markdown file and a pages folder, matching the same full-document plus page-level output pattern seen in the h — mistral-ai-financial-pdf-folder-structure-2.png
Input artifact: Input artifact (PDF document): 18-page table-heavy financial report. — llamaparse-sumitomo-financial-pdf-1.pdf
Output artifact: Output artifact (Image): The financial-report export contained a top-level markdown file and a pages folder, matching the same full-document plus page-level output pattern seen in the h — mistral-ai-financial-pdf-folder-structure-2.png
What changed: PDF document transformed into Image
Why it matters / Conclusion: If you need a hosted API that reliably gives you downloadable markdown with page-level inspection artifacts, Mistral AI delivered that consistently in this research.
Mistral AI returned markdown as downloadable output rather than a UI-only preview. In the hybrid earnings report, table-heavy financial report, and scanned research paper tests, the export pattern included a full-document markdown file plus page-level outputs, which makes it easier to inspect one page at a time or ingest the whole document at once.
Inside a page-level folder, the export included a markdown file plus separate image assets, showing that Mistral packages parsed content page by page instead of collapsing everything into one file.
The financial-report export contained a top-level markdown file and a pages folder, matching the same full-document plus page-level output pattern seen in the hybrid test.
Document hierarchy and reading order preservationUsually readable, but not consistently faithful to semantic structure.▾
Feature tested: Document hierarchy and reading order preservation
Result: Partial
Verdict: Usually readable, but not consistently faithful to semantic structure.
Expected behavior: Mistral AI generally kept headings attached to their paragraphs and preserved reading flow across native and scanned pages. The capability was exercised on a hybrid annual report section, a long narrative page from the financial report, a scanned multi-column research-paper section, a visually rich annual-report page, and the financial report's table of contents.
Test case: Image → Image
Input type: Image
Input used: Input artifact (Image): Target annual-report page with four section headings: Competition, Intellectual Property, Geographic Information, and Available Information, each followed by a — hybrid_earningspdf_sections_and_text.png
Observed output: Output artifact (Image): Mistral preserved the four section headings as distinct blocks with their associated paragraphs underneath, keeping the main reading order intact on this hybrid — mistral-ai-parsed-document-hierarchy.png
Input artifact: Input artifact (Image): Target annual-report page with four section headings: Competition, Intellectual Property, Geographic Information, and Available Information, each followed by a — hybrid_earningspdf_sections_and_text.png
Output artifact: Output artifact (Image): Mistral preserved the four section headings as distinct blocks with their associated paragraphs underneath, keeping the main reading order intact on this hybrid — mistral-ai-parsed-document-hierarchy.png
What changed: Image transformed into Image
Test case: Image → Image
Input type: Image
Input used: Input artifact (Image): Source page titled 'I. Summary of Operating Performance' with dense narrative paragraphs. — mistral-ai-financial-pdf-page-6-summary-operating-performance.png
Observed output: Output artifact (Image): The parsed output kept the section title, subsection title, and paragraph order readable, so this narrative page remained usable as structured markdown text. — mistral-ai-financial-pdf-parsed-operating-performance.png
Input artifact: Input artifact (Image): Source page titled 'I. Summary of Operating Performance' with dense narrative paragraphs. — mistral-ai-financial-pdf-page-6-summary-operating-performance.png
Output artifact: Output artifact (Image): The parsed output kept the section title, subsection title, and paragraph order readable, so this narrative page remained usable as structured markdown text. — mistral-ai-financial-pdf-parsed-operating-performance.png
What changed: Image transformed into Image
Test case: Image → Image
Input type: Image
Input used: Input artifact (Image): Scanned two-column research-paper section headed 'STAND PRESCRIPTIONS' with explanatory prose, numbered items, and subpoints. — scanned_pdf_standprescriptions_section.png
Observed output: Output artifact (Image): Despite the scanned multi-column source, Mistral kept the section heading attached to the explanatory paragraphs and preserved the numbered prescription flow in — mistral-ai-parsed-stand-prescriptions-hierarchy.png
Input artifact: Input artifact (Image): Scanned two-column research-paper section headed 'STAND PRESCRIPTIONS' with explanatory prose, numbered items, and subpoints. — scanned_pdf_standprescriptions_section.png
Output artifact: Output artifact (Image): Despite the scanned multi-column source, Mistral kept the section heading attached to the explanatory paragraphs and preserved the numbered prescription flow in — mistral-ai-parsed-stand-prescriptions-hierarchy.png
What changed: Image transformed into Image
Test case: Image → Image
Input type: Image
Input used: Input artifact (Image): Hybrid-report page titled 'A Growth Story Again' with title text, body copy, and a bullet-like list of growth outcomes. — hybrid_earningspdf_page.png
Observed output: Output artifact (Image): The content was recovered, but the hierarchy was only partially reconstructed: the page title and paragraphs survived, while deeper structure was flattened comp — mistral-ai-target-annual-report-markdown-viewer.png
Input artifact: Input artifact (Image): Hybrid-report page titled 'A Growth Story Again' with title text, body copy, and a bullet-like list of growth outcomes. — hybrid_earningspdf_page.png
Output artifact: Output artifact (Image): The content was recovered, but the hierarchy was only partially reconstructed: the page title and paragraphs survived, while deeper structure was flattened comp — mistral-ai-target-annual-report-markdown-viewer.png
What changed: Image transformed into Image
Test case: Image → Image
Input type: Image
Input used: Input artifact (Image): Financial report table of contents with Roman-numeral sections, nested subsections, and page numbers. — financialpdf_toc.png
Observed output: Output artifact (Image): The TOC entries were extracted as text, but the original navigational nesting was flattened, so the output no longer preserved a clear table-of-contents hierarc — mistral-ai-supplementary-materials-table-of-contents-2.png
Input artifact: Input artifact (Image): Financial report table of contents with Roman-numeral sections, nested subsections, and page numbers. — financialpdf_toc.png
Output artifact: Output artifact (Image): The TOC entries were extracted as text, but the original navigational nesting was flattened, so the output no longer preserved a clear table-of-contents hierarc — mistral-ai-supplementary-materials-table-of-contents-2.png
What changed: Image transformed into Image
Why it matters / Conclusion: Mistral AI is good at keeping pages readable and in order, but it is not the best choice when exact heading hierarchy and navigation structure need to survive conversion.
Mistral AI generally kept headings attached to their paragraphs and preserved reading flow across native and scanned pages. The capability was exercised on a hybrid annual report section, a long narrative page from the financial report, a scanned multi-column research-paper section, a visually rich annual-report page, and the financial report's table of contents.
Target annual-report page with four section headings: Competition, Intellectual Property, Geographic Information, and Available Information, each followed by a paragraph.
Mistral preserved the four section headings as distinct blocks with their associated paragraphs underneath, keeping the main reading order intact on this hybrid report page.
Source page titled 'I. Summary of Operating Performance' with dense narrative paragraphs.
The parsed output kept the section title, subsection title, and paragraph order readable, so this narrative page remained usable as structured markdown text.
Scanned two-column research-paper section headed 'STAND PRESCRIPTIONS' with explanatory prose, numbered items, and subpoints.
Despite the scanned multi-column source, Mistral kept the section heading attached to the explanatory paragraphs and preserved the numbered prescription flow into readable text.
Hybrid-report page titled 'A Growth Story Again' with title text, body copy, and a bullet-like list of growth outcomes.
The content was recovered, but the hierarchy was only partially reconstructed: the page title and paragraphs survived, while deeper structure was flattened compared with the source layout.
Financial report table of contents with Roman-numeral sections, nested subsections, and page numbers.
The TOC entries were extracted as text, but the original navigational nesting was flattened, so the output no longer preserved a clear table-of-contents hierarchy.
Table reconstructionGood on regular tables; weaker on nested header semantics.▾
Feature tested: Table reconstruction
Result: Partial
Verdict: Good on regular tables; weaker on nested header semantics.
Expected behavior: Mistral AI reconstructed several tables into usable markdown-like layouts, especially when the row and column logic was straightforward. It was tested on a Target financial summary table, a segment comparison table from the financial report, a more complex multi-level segment table, and a scanned results table with before/after measurement columns.
Test case: Image → Image
Input type: Image
Input used: Input artifact (Image): Target 2015 Annual Report 'Financial Summary' table with year columns 2015 through 2011 and multiple financial-result rows. — landing-ai-target-annual-report-financial-summary-table-2.png
Observed output: Output artifact (Image): Mistral reconstructed the financial summary into a readable table with the year columns and row labels aligned, preserving the relationship between metrics like — mistralai_hybrid_earnings_pdf_parsed_table.png
Input artifact: Input artifact (Image): Target 2015 Annual Report 'Financial Summary' table with year columns 2015 through 2011 and multiple financial-result rows. — landing-ai-target-annual-report-financial-summary-table-2.png
Output artifact: Output artifact (Image): Mistral reconstructed the financial summary into a readable table with the year columns and row labels aligned, preserving the relationship between metrics like — mistralai_hybrid_earnings_pdf_parsed_table.png
What changed: Image transformed into Image
Test case: Image → Image
Input type: Image
Input used: Input artifact (Image): Segment 'Orders Received' table with prior and present quarter values plus year-over-year change columns. — landing-ai-segment-results-table-2025-first-quarter.png
Observed output: Output artifact (Image): The extracted table retained the row labels, period labels, and change columns, making the segment comparison readable without manual re-keying. — mistral-ai-operating-cash-flow-comparison-quarterly-table.png
Input artifact: Input artifact (Image): Segment 'Orders Received' table with prior and present quarter values plus year-over-year change columns. — landing-ai-segment-results-table-2025-first-quarter.png
Output artifact: Output artifact (Image): The extracted table retained the row labels, period labels, and change columns, making the segment comparison readable without manual re-keying. — mistral-ai-operating-cash-flow-comparison-quarterly-table.png
What changed: Image transformed into Image
Test case: Image → Image
Input type: Image
Input used: Input artifact (Image): Financial segment table with nested header levels, subtotal and total columns, and footnote-marked columns E and F. — mistral-ai-financial-segment-table-millions-yen.png
Observed output: Output artifact (Image): Mistral preserved the numbers, but it merged distinct header levels into a flatter structure, weakening the parent-child relationships needed to interpret the c — mistral-ai-parsed-financial-segment-table-dark.png
Input artifact: Input artifact (Image): Financial segment table with nested header levels, subtotal and total columns, and footnote-marked columns E and F. — mistral-ai-financial-segment-table-millions-yen.png
Output artifact: Output artifact (Image): Mistral preserved the numbers, but it merged distinct header levels into a flatter structure, weakening the parent-child relationships needed to interpret the c — mistral-ai-parsed-financial-segment-table-dark.png
What changed: Image transformed into Image
Test case: Image → Image
Input type: Image
Input used: Input artifact (Image): Scanned results table comparing original diameter and diameter after harvest across four treatments, with inches and centimeters subcolumns. — mistral-ai-scanned-treatment-diameter-table.png
Observed output: Output artifact (Image): On this simpler scanned grid, Mistral kept the treatment rows and the inches/centimeters before-and-after measurements in a usable tabular layout, with only sur — mistral-ai-parsed-results-diameter-table.png
Input artifact: Input artifact (Image): Scanned results table comparing original diameter and diameter after harvest across four treatments, with inches and centimeters subcolumns. — mistral-ai-scanned-treatment-diameter-table.png
Output artifact: Output artifact (Image): On this simpler scanned grid, Mistral kept the treatment rows and the inches/centimeters before-and-after measurements in a usable tabular layout, with only sur — mistral-ai-parsed-results-diameter-table.png
What changed: Image transformed into Image
Why it matters / Conclusion: For regular financial tables and simpler scanned grids, Mistral AI was usable. For tables where nested headers carry important meaning, fidelity dropped.
Mistral AI reconstructed several tables into usable markdown-like layouts, especially when the row and column logic was straightforward. It was tested on a Target financial summary table, a segment comparison table from the financial report, a more complex multi-level segment table, and a scanned results table with before/after measurement columns.
Target 2015 Annual Report 'Financial Summary' table with year columns 2015 through 2011 and multiple financial-result rows.
Mistral reconstructed the financial summary into a readable table with the year columns and row labels aligned, preserving the relationship between metrics like Sales, EBIT, and Net earnings and their values.
Segment 'Orders Received' table with prior and present quarter values plus year-over-year change columns.
The extracted table retained the row labels, period labels, and change columns, making the segment comparison readable without manual re-keying.
Financial segment table with nested header levels, subtotal and total columns, and footnote-marked columns E and F.
Mistral preserved the numbers, but it merged distinct header levels into a flatter structure, weakening the parent-child relationships needed to interpret the column semantics precisely.
Scanned results table comparing original diameter and diameter after harvest across four treatments, with inches and centimeters subcolumns.
On this simpler scanned grid, Mistral kept the treatment rows and the inches/centimeters before-and-after measurements in a usable tabular layout, with only surrounding narrative added above and below the table.
Visual asset retentionVisuals were retained as page-linked assets instead of being dropped.▾
Feature tested: Visual asset retention
Result: Passed
Verdict: Visuals were retained as page-linked assets instead of being dropped.
Expected behavior: Mistral AI extracted non-text document elements into page-specific assets and kept them associated with the markdown workflow. This was tested on the hybrid earnings report, which included charts and a scanned signature/stamp region, and on the scanned research paper, which included chart content referenced from the surrounding text.
Test case: PDF document → Image
Input type: PDF document
Input used: Input artifact (PDF document): Hybrid earnings report containing charts and a scanned signature or stamp region. — llamaparse-hybrid-earnings-pdf-1.pdf
Observed output: Output artifact (Image): The page-level export for the hybrid report included separate image assets alongside markdown, showing that visuals were not dropped during conversion. — mistral-ai-windows-explorer-page-folder.png
Input artifact: Input artifact (PDF document): Hybrid earnings report containing charts and a scanned signature or stamp region. — llamaparse-hybrid-earnings-pdf-1.pdf
Output artifact: Output artifact (Image): The page-level export for the hybrid report included separate image assets alongside markdown, showing that visuals were not dropped during conversion. — mistral-ai-windows-explorer-page-folder.png
What changed: PDF document transformed into Image
Test case: PDF document → Image
Input type: PDF document
Input used: Input artifact (PDF document): Hybrid earnings report markdown view with embedded page assets. — llamaparse-hybrid-earnings-pdf-1.pdf
Observed output: Output artifact (Image): The markdown workflow surfaced page content together with extracted page assets, letting report visuals remain associated with the original page rather than dis — mistral-ai-vscode-markdown-embedded-assets.png
Input artifact: Input artifact (PDF document): Hybrid earnings report markdown view with embedded page assets. — llamaparse-hybrid-earnings-pdf-1.pdf
Output artifact: Output artifact (Image): The markdown workflow surfaced page content together with extracted page assets, letting report visuals remain associated with the original page rather than dis — mistral-ai-vscode-markdown-embedded-assets.png
What changed: PDF document transformed into Image
Test case: Image → Image
Input type: Image
Input used: Input artifact (Image): Scanned research-paper page containing a chart referenced from the DISCUSSION section. — scanned-pdf-grouped-chart.png
Observed output: Output artifact (Image): The parsed output included an extracted image file and a markdown file in the same workspace, indicating that chart images were retained and linked to their sou — mistral-ai-editor-discussion-with-embedded-chart.png
Input artifact: Input artifact (Image): Scanned research-paper page containing a chart referenced from the DISCUSSION section. — scanned-pdf-grouped-chart.png
Output artifact: Output artifact (Image): The parsed output included an extracted image file and a markdown file in the same workspace, indicating that chart images were retained and linked to their sou — mistral-ai-editor-discussion-with-embedded-chart.png
What changed: Image transformed into Image
Test case: Image → Image
Input type: Image
Input used: Input artifact (Image): Blurry footer stamp or signature region from the hybrid earnings report. — blurry_marker.png
Observed output: Output artifact (Image): Mistral recovered the footer stamp text, including 'Minneapolis, Minnesota,' the date, the firm name, and the page number, showing usable OCR even on a blurry f — mistral-ai-document-footer-stamp.png
Input artifact: Input artifact (Image): Blurry footer stamp or signature region from the hybrid earnings report. — blurry_marker.png
Output artifact: Output artifact (Image): Mistral recovered the footer stamp text, including 'Minneapolis, Minnesota,' the date, the firm name, and the page number, showing usable OCR even on a blurry f — mistral-ai-document-footer-stamp.png
What changed: Image transformed into Image
Why it matters / Conclusion: The research supports Mistral AI as a good choice when charts, images, and scanned visual regions need to stay attached to the markdown output instead of being silently omitted.
Mistral AI extracted non-text document elements into page-specific assets and kept them associated with the markdown workflow. This was tested on the hybrid earnings report, which included charts and a scanned signature/stamp region, and on the scanned research paper, which included chart content referenced from the surrounding text.
The page-level export for the hybrid report included separate image assets alongside markdown, showing that visuals were not dropped during conversion.
The markdown workflow surfaced page content together with extracted page assets, letting report visuals remain associated with the original page rather than disappearing into plain text.
Scanned research-paper page containing a chart referenced from the DISCUSSION section.
The parsed output included an extracted image file and a markdown file in the same workspace, indicating that chart images were retained and linked to their source page.
Blurry footer stamp or signature region from the hybrid earnings report.
Mistral recovered the footer stamp text, including 'Minneapolis, Minnesota,' the date, the firm name, and the page number, showing usable OCR even on a blurry footer region.
Pricing & Access
Is This Right For You?
A side-by-side guide based on our hands-on testing.
Use Case Track Record
Related Pages
Banner Preview
How the embed badge will look on your site
Embed HTML
Copy this code to your website source
Quick Integration Guide
- 1Copy the HTML code block above.
- 2Paste it into your site's HTML or CMS editor.
- 3Banner appears instantly on your page.
- 4Links back to your tool profile here.
Similar Tools
Discover more AI tools like Mistral AI to enhance your workflow.