BlazePDF

A benchmark-driven Rust PDF reader core, CLI, and lightweight desktop viewport. Classification, first-page text, native rendering, and full Markdown extraction are separate cost tiers: opening a document need not pay for every capability at once.

Latest snapshot: 40 gold medals across 40 benchmark cards. Eight documents × five workloads, with nine measured samples per BlazePDF card and one excluded warmup. This describes this corpus and machine—not every PDF, feature set, or computer.

Quick start

After building the CLI below, inspect your PDF without extracting Markdown:

.\target\release\blazepdf.exe --inspect document.pdf

Example output format (values depend on the document and machine):

pages: 12
kind: Text
signed: true
processed: 37 ms

This example is the exact output asserted by the CLI formatting test, not a new benchmark measurement. signed indicates a signature marker, not validation.

Benchmark methodology · Contributing · Security · Changelog · Release process

The release badge links to published releases; until the first release exists, build from source. It does not imply downloadable binaries are already available.

Capabilities

• Classification and page count without building Markdown.
• First-page text layers with rotation information and bounded interactive caching.
• Image/ambiguous-page routing to raster without blocking first paint on OCR.
• First-page rendering at 150 DPI through a separately supplied PDFium runtime.
• Structured Markdown extraction and a minimal egui desktop viewport.

The signature marker is routing metadata, not cryptographic signature validation. BlazePDF is experimental, not a feature-equivalent replacement for commercial editors. Output-equality checks compare before/after BlazePDF builds; the benchmark does not establish identical rendering quality across different products.

Capability	BlazePDF status	Scope / limitation
Classification and page count	Implemented	Separate low-cost inspection path
First-page text	Implemented	Rotation information and bounded cache
Markdown extraction	Implemented	Document-dependent fidelity; not editing
Native first-page raster	Implemented	Requires separately supplied PDFium
Desktop viewing	Experimental	Minimal viewport, not a full editor
Signature detection	Marker only	No cryptographic verification
OCR, editing, annotation	Not promised	Not established by these benchmarks

This is a capability comparison with full-featured PDF products, not an assertion that another named product lacks any feature.

Platform support

Platform	Status	Evidence
Windows x64 / MSVC	Locally tested	Published tests and benchmark snapshot; Windows CI configured
Linux	Unverified	No published build, test or benchmark evidence
macOS	Unverified	No published build, test or benchmark evidence

CI covers the non-GUI CLI and selected tests, not PDFium rendering or desktop launch performance. Its acceptance command explicitly skips the native-render test, which requires PDFium. Toolchain badges record the local snapshot, not CI minimums.

Build and run

Measured environment: Windows 11, Intel Core i7-12700H (20 logical CPUs), 31.7 GB RAM, Rust 1.98.0, Python 3.13.2, release builds. The benchmark harness uses Windows-specific executables/window detection; other platforms were not measured in this snapshot.

Install Rust and the MSVC C++ build tools on Windows, then:

git clone https://github.com/marksmayo/blazepdf.git
cd blazepdf
cargo build --release --locked

For the library and CLI without the GUI feature:

cargo build --release --locked --no-default-features --bin blazepdf

.\target\release\blazepdf.exe --inspect document.pdf
.\target\release\blazepdf.exe --first-page document.pdf
.\target\release\blazepdf.exe document.pdf
.\target\release\blazepdf.exe --render document.pdf page.ppm
.\target\release\blazepdf-window.exe document.pdf

Native rendering and viewport raster fallback require a compatible PDFium library. Set PDFIUM_LIB_PATH to its library path if automatic discovery cannot find it. PDFium and other vendors' installers are not bundled.

Benchmark methodology and limitations

The chart is a fixed snapshot of classification medians in milliseconds (lower is better), not a cross-workload average. See the tables for exact values and coverage.

BlazePDF's latest measurements are from 2 October 2026 (NZDT). Competitor rows retain the 30 September 2026 reference. This is a mixed-date snapshot, not a fresh simultaneous rerun of every competitor. Other applications were running during development, and launch timings vary. Competitors have five or six samples per successful row; BlazePDF has nine. All retained samples are in the raw JSON.

Only identical document/workload pairs are compared. Classification, first-page text, Markdown extraction, 150-DPI rendering and launch-to-visible-window stay separate. Missing/failed measurements are not zero or losses. BlazePDF's matrix uses one complete run, not a collection of historical per-card bests.

The following scores and timings are generated directly from canonical evidence.

Medal / score table

Source: benchmark-20261002T072651Z-merged.json. 40 document/workload cards.

Score is average rank over successfully measured cards (lower is better), not an average of incompatible latencies. Missing cards are ignored: read score alongside coverage. Ties share ranks (1, 1, 3).

Product	Gold	Silver	Bronze	Other placements	Coverage	Score ↓
BlazePDF	40	0	0	0	40 / 40	1.00
MuPDF	0	30	5	5	40 / 40	2.45
PDF24 Reader	0	1	7	0	8 / 40	2.88
pdf-inspector	0	8	8	7	23 / 40	3.04
PDF24 bundled qpdf	0	0	2	6	8 / 40	4.38
SumatraPDF	0	1	9	14	24 / 40	4.54
Google Chrome PDF viewer	0	0	0	8	8 / 40	4.62
Okular	0	0	0	6	6 / 40	4.67
CalyPdf	0	0	1	23	24 / 40	4.96
Foxit PDF Reader	0	0	0	8	8 / 40	6.00
Sejda PDF Desktop	0	0	0	8	8 / 40	6.62
Xodo PDF Reader	0	0	0	8	8 / 40	8.75

Catalogued but not measured/scored: Adobe Acrobat Reader.

Timing comparisons

Median milliseconds, including the declared adapter's launch cost. Reference is the fastest measured non-BlazePDF product for the same document/workload. Gap = reference − BlazePDF; ratio = reference / BlazePDF. Positive gaps and ratios above 1 favor BlazePDF.

Full timing tables for every measured product.

classify + extract markdown

Document	BlazePDF (ms)	Fastest reference	Reference (ms)	Gap (ms)	Ratio
Financial table	22.112	MuPDF	28.259	+6.147	1.28×
Image document	270.299	MuPDF	298.699	+28.400	1.11×
Multi-column	26.242	pdf-inspector	69.790	+43.548	2.66×
Overlapping text	18.367	MuPDF	80.300	+61.933	4.37×
RTL Hebrew	16.577	pdf-inspector	34.018	+17.441	2.05×
Rotation	16.893	MuPDF	35.139	+18.246	2.08×
Signed PDF	20.909	MuPDF	46.613	+25.704	2.23×
Text-heavy	31.393	MuPDF	38.074	+6.681	1.21×

classify only

Document	BlazePDF (ms)	Fastest reference	Reference (ms)	Gap (ms)	Ratio
Financial table	12.381	pdf-inspector	22.855	+10.474	1.85×
Image document	21.419	MuPDF	23.506	+2.087	1.10×
Multi-column	14.033	MuPDF	20.671	+6.638	1.47×
Overlapping text	18.301	MuPDF	24.308	+6.007	1.33×
RTL Hebrew	14.543	pdf-inspector	27.489	+12.946	1.89×
Rotation	13.050	MuPDF	21.672	+8.622	1.66×
Signed PDF	16.811	MuPDF	29.015	+12.204	1.73×
Text-heavy	13.455	MuPDF	14.389	+0.934	1.07×

launch to visible document window

Document	BlazePDF (ms)	Fastest reference	Reference (ms)	Gap (ms)	Ratio
Financial table	20.593	MuPDF	115.028	+94.435	5.59×
Image document	20.455	PDF24 Reader	372.615	+352.160	18.22×
Multi-column	20.254	MuPDF	90.327	+70.073	4.46×
Overlapping text	20.287	MuPDF	128.300	+108.013	6.32×
RTL Hebrew	21.606	MuPDF	44.140	+22.534	2.04×
Rotation	20.355	MuPDF	72.137	+51.782	3.54×
Signed PDF	20.033	MuPDF	72.812	+52.779	3.63×
Text-heavy	19.940	MuPDF	67.089	+47.149	3.36×

open document + first-page text layer

Document	BlazePDF (ms)	Fastest reference	Reference (ms)	Gap (ms)	Ratio
Financial table	16.066	pdf-inspector	40.013	+23.947	2.49×
Image document	190.402	SumatraPDF	259.716	+69.314	1.36×
Multi-column	16.149	pdf-inspector	23.606	+7.457	1.46×
Overlapping text	19.844	pdf-inspector	47.337	+27.493	2.39×
RTL Hebrew	13.055	MuPDF	25.190	+12.135	1.93×
Rotation	12.890	pdf-inspector	39.541	+26.651	3.07×
Signed PDF	17.627	MuPDF	39.021	+21.394	2.21×
Text-heavy	24.715	MuPDF	56.882	+32.167	2.30×

render first page at 150 DPI

Document	BlazePDF (ms)	Fastest reference	Reference (ms)	Gap (ms)	Ratio
Financial table	43.791	MuPDF	107.538	+63.747	2.46×
Image document	634.231	MuPDF	2504.587	+1870.356	3.95×
Multi-column	32.257	MuPDF	143.759	+111.502	4.46×
Overlapping text	126.200	MuPDF	300.778	+174.578	2.38×
RTL Hebrew	30.395	MuPDF	72.218	+41.823	2.38×
Rotation	35.906	MuPDF	154.254	+118.348	4.30×
Signed PDF	35.599	MuPDF	131.809	+96.210	3.70×
Text-heavy	46.059	MuPDF	143.413	+97.354	3.11×

Reports and evidence

• Interactive workload graphs and medal/score board.
• Interactive dashboard and distribution statistics.
• All-products timing tables.
• Latest 40 BlazePDF measurements.
• Complete comparison snapshot.
• Unchanged competitor reference.
• Ranked optimization candidates and implementation evidence.

GitHub displays HTML source rather than executing these reports. Download either HTML file and open it locally for the interactive view; Markdown tables work on GitHub.

Reproduce the corpus, tests and benchmarks

Third-party PDFs are not redistributed. Download the eight commit-pinned corpus files and verify their hashes before running acceptance tests or benchmarks:

python scripts/bootstrap_benchmark_corpus.py
python scripts/bootstrap_benchmark_corpus.py --check
cargo test --no-default-features --bin blazepdf --test reader_acceptance
cargo test --no-default-features --test inspection_allocation_budget --test extraction_allocation_budget
python scripts/test_workload_graphs.py
python scripts/test_update_canonical_reports.py
python scripts/test_readme_benchmarks.py
python scripts/update_readme_benchmarks.py --check

Our modified pdf-inspector is preserved as ordinary source: no submodule setup is required. The eight-file bootstrap supplies the top-level acceptance suite. Its larger internal regression suite needs additional upstream fixtures not provisioned by this bootstrap.

With PDFium configured, rerun the full five-workload BlazePDF matrix:

python scripts/benchmark.py --reader blazepdf-first-page --reader blazepdf-inspect --reader blazepdf-render --reader blazepdf --reader blazepdf-window --repeats 9 --no-canonical

To compare the whole new run with the retained reference, replace <new-run> with its timestamped filename:

python scripts/merge_benchmark_runs.py results/benchmark-20260930T040051Z-merged.json results/<new-run>.json results/<new-run>-merged.json
python scripts/update_canonical_reports.py
python scripts/update_readme_benchmarks.py

For a fresh all-product comparison, install/configure the readers in benchmarks/readers.json, then run python scripts/benchmark.py. Executable paths in that manifest describe the original machine: update them for your installation. Missing installations are recorded as skipped. Other vendors' software is not bundled.

Layout

Path	Purpose
src/	Reader core, CLI and viewport
crates/pdf-inspector/	Modified PDF core and bundled CMaps
tests/	Acceptance and allocation regression gates
benchmarks/	Corpus, pinned fixture sources and reader/product configuration
scripts/	Benchmarks, comparisons and report/documentation generation
results/	Published final raw evidence and standalone reports
docs/benchmark-timings.md	Timings for every measured product

Repository line counts

The line-count badges are a snapshot of physical lines in Git-tracked .rs, .py and .html files, including comments and blank lines. Rust includes the modified vendored PDF core and tests (83 files); Python covers the harness, generators and tests (21 files). HTML covers the two generated, minified benchmark reports (11 physical lines), not HTML templates embedded in Python strings. These are file-line counts, not executable-code counts or live CI metrics. Untracked vendor checkouts, build outputs and downloaded fixtures are excluded.

License and attribution

BlazePDF's own code is MIT licensed. The Firecrawl PDF core, Adobe CMaps, downloaded documents and runtime dependencies retain their notices and terms. See NOTICE.md for source provenance and attribution.

Verification

The local publication checks passed on 2 October 2026: three Rust CLI tests, 18 reader acceptance tests, both allocation regression gates, three Python documentation/fixture tests, workload-graph checks, and both canonical-report selection tests. The README timing tables also match the canonical JSON, and all eight commit-pinned upstream fixture hashes were verified.

That is a verified snapshot, not a claim that every optional test in the repository has run. The CI badge separately reports the workflow's actual status. Rust and Python badges show the tested toolchain versions, not minimum supported versions. The gold badge describes the published benchmark snapshot and its methodology above.

Contributions are welcome: open an issue to discuss a change, or submit a pull request with focused tests. If this project is useful to you, click Star at the top of the repository to support it.