ZeRechnung
Language: English

Benchmarks

How fast is it?

ZeRechnung was measured against 14 other Go libraries, then against 32 TypeScript, Python and Rust libraries: same files from the official ZUGFeRD corpus, same machine, same harness. Every result is checked before it is timed.

Times of ZeRechnung

Extract the XML

30.2 µs

×2.5 faster than the next Go library

Read the XML into a model

16.9 µs

×7.3 faster than the next Go library

Read and validate

24.3 µs

×5.1 faster than the next Go library

Write the XML

3.0 µs

×6.7 faster than the next Go library

Embed the XML in a PDF

89.2 µs

×5.9 faster than the next Go library

Check a whole Factur-X PDF

74.2 µs

×15 faster than the next Go library

Measured on 17 November 2026 on an Apple M4 Pro, on one goroutine, with ZeRechnung v0.1.0.

Against the other Go libraries

The other libraries read XML with encoding/xml, whose tokenizer alone takes 100 µs on a 13 KB invoice. ZeRechnung has its own XML reader and goes well below that floor. A lazy PDF reader, which reads only the objects it needs, and an incremental update to add the invoice do the rest.

The bars are on a logarithmic scale: each grid line is ten times slower than the one before. Times are medians of three runs.

Extract the XML

EN 16931 PDF, 149 KB
LibraryTime, logarithmic scaleTime per operation
ZeRechnung 30.2 µs
speedata 75.2 µs ×2.5
litho 106 µs ×3.5
papierleicht 113 µs ×3.7
DCNWS 209 µs ×6.9
pdfcpu 317 µs ×10
fX 332 µs ×11
dotwave 343 µs ×11
angelo 356 µs ×12
kanjo 386 µs ×13
gopdfattach 396 µs ×13
andee 516 µs ×17
taha fails: it looks for the XML with a regular expression, which misses compressed streams

Read the XML into a model

CII XML, 13 KB
LibraryTime, logarithmic scaleTime per operation
ZeRechnung 16.9 µs
fXloses the VAT breakdown 124 µs ×7.3
angelo 131 µs ×7.8
dotwave 144 µs ×8.5
tahastores amounts as whole cents 147 µs ×8.7
DCNWS 165 µs ×9.8
andee 182 µs ×11
speedata 225 µs ×13
kanjo 253 µs ×15
gobl.de 564 µs ×33

Dashed line: 100 µs, the time encoding/xml takes just to split the file into tokens.

Read and validate

CII XML, 13 KB
LibraryTime, logarithmic scaleTime per operationValid invoices reported as invalid
ZeRechnung 24.3 µs 6/26
fX 123 µs ×5.1 26/26
pdf0 126 µs ×5.2 6/26
taha 147 µs ×6.0 15/26
angelo 172 µs ×7.1 4/26
andee 208 µs ×8.6 22/26
kanjo 293 µs ×12 9/26
DCNWS 754 µs ×31 2/26
speedata 1.28 ms ×53 0/26
gobl.de 1.77 ms ×73 20/26
papierleicht 14.65 ms ×603 6/26

Last column: how many of the 26 valid invoices of the corpus the library reports as invalid. The official schematrons report 6, so 6 is the right answer.

Write the XML

CII XML, 13 KB
LibraryTime, logarithmic scaleTime per operation
ZeRechnung 3.0 µs
fX 20.2 µs ×6.7
dotwave 22.1 µs ×7.4
speedata 30.5 µs ×10
DCNWS 36.2 µs ×12
kanjo 98.3 µs ×33
andee 311 µs ×104
taha 652 µs ×217

Embed the XML in a PDF

host PDF with xref streams, 144 KB
LibraryTime, logarithmic scaleTime per operation
ZeRechnung 89.2 µs
pdf0 522 µs ×5.9
pdfcpulow-level recipe written for the benchmark 546 µs ×6.1
gopdfattach 1.12 ms ×13
dotwave 1.24 ms ×14
DCNWS 1.66 ms ×19
lithoconverts and rewrites the whole PDF 36.16 ms ×405
kanjo returns a PDF that cannot be read back
andee returns a PDF that cannot be read back

Check a whole Factur-X PDF

EN 16931 PDF, 149 KB
LibraryTime, logarithmic scaleTime per operation
ZeRechnung 74.2 µs
DCNWS 1.14 ms ×15
pdf0 2.20 ms ×30
papierleicht 15.78 ms ×213
litho 19.87 ms ×268

Against other languages

For each operation, the fastest library of each language among those that succeed. ZeRechnung is the fastest in 13 columns out of 16. Rust comes closest, without applying all the EN 16931 rules. The exceptions are explained on the wiki.

Extract the XML

EN 16931 PDF, 149 KB
LibraryTime, logarithmic scaleTime per operation
ZeRechnungGo 29.3 µs
einvoice-validatorRust 32.5 µs ×1.1
speedataGo 75.2 µs ×2.6
zugferd-pdfPython 78.5 µs ×2.7
@attestwire/en16931TypeScript 184 µs ×6.3

Read the XML into a model

CII XML, 13 KB
LibraryTime, logarithmic scaleTime per operation
ZeRechnungGo 16.3 µs
zugferd-rsRust 24.0 µs ×1.5
@ekwo-ai/factur-xTypeScript 69.6 µs ×4.3
fXGo 124 µs ×7.6
pychevalPythonreads 7 of the 26 XML files 221 µs ×14

Read the XML into a model

CII XML, 1.3 MB, with a base64 attachment
LibraryTime, logarithmic scaleTime per operation
mcp-einvoicing-dePythonlxml, written in C, keeps the attachment as text 702 µs
ZeRechnungGo 1.11 ms ×1.6
pdf-invoiceRust 1.63 ms ×2.3
@conformo/toolkitTypeScript 1.74 ms ×2.5
dotwaveGo 6.76 ms ×9.6

Validate with all EN 16931 rules

CII XML, 13 KB
LibraryTime, logarithmic scaleTime per operation
ZeRechnungGo 23.1 µs
pdf0Go 126 µs ×5.5
einvoice-validatorRust 473 µs ×20
@sweberdev/summandTypeScript 802 µs ×35
saxonchePythonwith the official XSLT 903 µs ×39

Write the XML

CII XML, 13 KB
LibraryTime, logarithmic scaleTime per operation
ZeRechnungGo 3.1 µs
zugferd-rsRust 6.7 µs ×2.2
@ekwo-ai/factur-xTypeScriptmodel built by hand 10.9 µs ×3.5
fXGo 20.2 µs ×6.5
pychevalPython 76.6 µs ×25

Embed the XML in a PDF

host PDF with xref streams, 144 KB
LibraryTime, logarithmic scaleTime per operation
ZeRechnungGo 87.8 µs
pdf-invoiceRust 178 µs ×2.0
pdf0Go 522 µs ×5.9
mcp-einvoicing-dePython 662 µs ×7.5
@stackforge-eu/factur-xTypeScript 1.14 ms ×13

Check a whole Factur-X PDF

EN 16931 PDF, 149 KB
LibraryTime, logarithmic scaleTime per operation
ZeRechnungGo 75.0 µs
zugferd-pdfPythonchecks the container only 212 µs ×2.8
@attestwire/en16931TypeScript 359 µs ×4.8
en16931Rust 479 µs ×6.4
DCNWSGo 1.14 ms ×15

Fast, and right

Speed means nothing if the verdicts are wrong. On the 26 valid invoices of the official corpus, the official schematrons report 6: rule BR-CL-25 twice, and CII-SR-465 on the four XRECHNUNG examples. ZeRechnung reports exactly these 6, and extracts the XML of all 104 PDF files of the corpus. Its CI also compares every pull request with the KoSIT and Mustang validators.

Among the 32 libraries in other languages, six give the verdicts of the reference. The most downloaded ones check only the XSD schema, or only generate invoices.

Method

One module per library, called through its public API in the most direct way its documentation offers. Every timing starts with a correctness check: the extracted XML must be a CII invoice, and a PDF produced by an injection must be read back. The harnesses are in the repository, so anyone can run them again.

MachineApple M4 Pro · 24 GB · macOS 26.6
Go1.26.4 · go test -bench . -benchmem -count 3 -benchtime 1s
TypeScriptNode 22.15.0
PythonCPython 3.13.18
Rustrustc 1.93.1, release
Value shownmedian of 3 runs, on one thread, with the bytes already in memory