PDF parsing has always been one of the hardest parts of web scraping. Most PDFs aren't simple text — they contain scanned pages, multi-column layouts, tables, formulas, and mixed content. Until now, every solution forced a tradeoff: fast but inaccurate, or accurate but too slow to run at scale. Today we're shipping Fire-PDF, a PDF parsing engine built to eliminate that tradeoff.

P.S. Every PDF sent through our API now goes through Fire-PDF automatically. No configuration needed.

Fire-PDF is our new Rust-based PDF parsing engine.

It converts any PDF — scanned, text-based, or mixed — into structured markdown. Our open-source Rust library pdf-inspector classifies each page in milliseconds. Text-based pages go straight to native extraction, skipping GPU entirely. Only scanned or image-heavy content hits the neural layout model and OCR.

The result is clean markdown with correct reading order, preserved tables, formulas in LaTeX, and proper multi-column structure.

Compared to our previous PDF parser, Fire-PDF is 3.5-5.7x faster — averaging under 400ms per page.

Speed comes from two places:

• First, text-based pages never touch GPU — they get native extraction through pdf-inspector in milliseconds.
• Second, the GPU fleet uses lane-based routing to isolate requests by document size, so a 200-page report never impacts latency for a single-page invoice.

Most documents are not fully scanned. A financial report might have 150 text-based pages and 60 scanned ones. With our previous pipeline, all of it went through OCR.

With Fire-PDF, only the pages that need it do.

pdf-inspector is our open-source Rust library that classifies every page by analyzing PDF internals — font encodings, text operators, and image coverage — in milliseconds, without rendering.

• Text-based pages get instant native extraction.
• Only scanned or image-heavy pages go through GPU.

For mixed documents, this can eliminate GPU processing for the majority of pages, which translates directly into faster processing and lower cost.

Speed alone isn't enough if tables come out garbled or multi-column text comes out of order. Fire-PDF uses a neural document layout model to detect text blocks, tables, formulas, images, headers, and footers individually — then handles each region type correctly.

• Tables get higher token limits and up to 25 seconds to generate accurate markdown table output
• Formulas get formula-specific prompts and are preserved in LaTeX
• Text regions get tight 12-second, 256-token budgets for efficiency
• Reading order is predicted neurally, with XY-cut projection as a fallback for multi-column layouts

Each region gets tuned parameters rather than one-size-fits-all OCR. The difference shows on documents that previously came through as jumbled text. Think: financial tables, academic papers with equations, legal filings with dense columns.

Fire-PDF runs every PDF through five stages:

1. Classify — pdf-inspector scans the PDF's internal structure in milliseconds, classifying each page as text-based or needing OCR
2. Render — Pages needing OCR are rendered to images at 200 DPI. Oversized pages are automatically capped or sliced
3. Layout Detection — Rendered images go through a neural document layout model on GPU, returning bounding boxes, element types, and reading order
4. Extraction — Text-based pages use native extraction (no GPU). Scanned regions are sent to a vision-language model (GLM-OCR) with task-specific prompts and parameters per region type
5. Assembly — Results are sorted by reading order and assembled into markdown. Tables become markdown tables. Formulas are preserved in LaTeX. Geometric deduplication removes overlapping detections

Fire-PDF is live for all Firecrawl users. Every PDF you send through the API uses it automatically - no configuration required.

Get started with Fire-PDF · Read the docs