Benchmarks

Takeaways

From the run below, on x86-64 Linux with Node.js 24:

• zstd is faster than gzip and brotli in comprs, for compression and decompression alike, with a better ratio than gzip. LZ4 compresses faster still on four of the six inputs, at a lower ratio, but decompresses faster than zstd only for 150 bytes of text and 84 KB of JSON. Against the zstd of node:zlib at the same level, comprs compresses 1.1–2.7 times and decompresses 1.4–3.6 times as fast, to the same size.
• gzip decompression is 3.5–16 times as fast as pako and 1.0–2.0 times as fast as node:zlib; fflate is faster on incompressible data and on 150 bytes. deflate (zlib) decompression is 1.5–5 times as fast as pako, and 0.8–1.3 times as fast as node:zlib.
• gzip and deflate compression of 84 KB of JSON is 1.5 times as fast as node:zlib and 4.2–4.7 times as fast as pako, at the same ratio within 1%. comprs compresses with zlib-rs, which follows the algorithms of zlib-ng: on extremely repetitive input, such as one sentence repeated, it trades ratio for speed (82:1 against 122:1 for zlib), so the speeds on those inputs are not like for like. For 150 bytes, node:zlib is faster.
• brotli comes from libbrotli, the C library of node:zlib: it compresses to the same size at about the same speed, and decompresses up to 1.5 times as fast.
• Node.js streams of comprs are 3.7–7.9 times as fast as those of node:zlib for 1 MB in 16 KB chunks.
• These numbers are from the native addon; the WebAssembly build was not measured.

Tables

Measured on 2026-10-25 with comprs 2.3.0 (fb6a968), Node.js v24.21.0, pako 3.0.2 and fflate 0.8.3, on AMD EPYC 7763 64-Core Processor × 4 (linux/x64), in this run of the Bench Report workflow, on a GitHub-hosted runner (ubuntu24, image 20261023.334.3). Settings, the same for every library in a table: gzip level 6 · deflate level 6 · brotli quality 6, lgwin 22 · brotli quality 11, lgwin 22 · zstd level 3 · streams: gzip level 6, zstd level 3 · cross-algorithm: each format’s default settings (lz4 has no levels). Speeds are MB/s of uncompressed data (1 MB = 1,000,000 bytes), from the median time of a call. Every library decompresses the output of comprs, so that only the decoders differ, except in the cross-algorithm table, where each format decompresses its own. The compression ratio is the input size divided by the output size.

Cross-algorithm comparison: zstd vs gzip vs brotli vs lz4 (comprs only, default settings)

Compression (MB/s of uncompressed data, higher is better)

Input	zstd	gzip	brotli	lz4
text 150B	56.5	13.2	15.5	85.5
JSON 84KB	583	233	76.9	1,122
text 45KB	3,535	1,715	1,602	3,194
random 10KB	932	80.8	108	1,672
random 1MB	3,811	49.9	381	2,891
patterned 1MB	6,274	2,866	1,331	3,703

Decompression (MB/s of uncompressed data, higher is better)

Input	zstd	gzip	brotli	lz4
text 150B	103	59.7	46.6	111
JSON 84KB	1,757	1,395	889	2,374
text 45KB	8,026	2,992	1,034	1,772
random 10KB	4,822	2,233	3,108	2,796
random 1MB	13,711	4,000	6,864	4,386
patterned 1MB	17,150	4,398	1,136	1,861

Compression ratio (input size ÷ output size, higher is better)

Input	zstd	gzip	brotli	lz4
text 150B	4.84	3.95	5.56	2.78
JSON 84KB	9.25	7.54	10.4	4.60
text 45KB	285	81.8	286	105
random 10KB	1.00	1.00	1.00	1.00
random 1MB	1.00	1.00	1.00	1.00
patterned 1MB	2,755	128	3,846	199

gzip, level 6: comprs vs pako vs fflate vs node:zlib

Compression (MB/s of uncompressed data, higher is better)

Input	comprs	pako	fflate	node:zlib
text 150B	12.4	2.16	6.98	19.7
JSON 84KB	231	49.0	36.3	152
text 45KB	1,709	84.5	83.7	294
random 10KB	81.2	32.5	36.5	97.9
random 1MB	49.7	20.4	36.0	48.7
patterned 1MB	2,834	77.7	83.2	291

Decompression (MB/s of uncompressed data, higher is better)

Input	comprs	pako	fflate	node:zlib
text 150B	59.2	17.1	189	59.0
JSON 84KB	1,385	191	375	943
text 45KB	3,001	242	803	1,522
random 10KB	2,184	300	6,698	1,845
random 1MB	3,987	365	9,362	2,282
patterned 1MB	4,265	265	883	2,286

Compression ratio (input size ÷ output size, higher is better)

Input	comprs	pako	fflate	node:zlib
text 150B	3.95	3.95	4.05	3.95
JSON 84KB	7.54	7.58	8.09	7.58
text 45KB	81.8	122	122	122
random 10KB	1.00	1.00	1.00	1.00
random 1MB	1.00	1.00	1.00	1.00
patterned 1MB	128	237	237	237

deflate, level 6: comprs vs pako vs fflate vs node:zlib

Compression (MB/s of uncompressed data, higher is better)

Input	comprs	pako	fflate	node:zlib
text 150B	13.1	2.20	7.48	20.7
JSON 84KB	236	56.7	41.2	158
text 45KB	2,022	113	107	324
random 10KB	83.7	38.2	40.6	99.2
random 1MB	50.0	22.0	40.7	49.5
patterned 1MB	3,708	96.2	111	332

Decompression (MB/s of uncompressed data, higher is better)

Input	comprs	pako	fflate	node:zlib
text 150B	77.6	17.4	2.19	65.1
JSON 84KB	1,521	365	259	1,194
text 45KB	3,284	608	358	2,211
random 10KB	3,007	952	2,527	3,431
random 1MB	6,015	4,064	3,256	5,018
patterned 1MB	4,066	790	664	5,096

Compression ratio (input size ÷ output size, higher is better)

Input	comprs	pako	fflate	node:zlib
text 150B	7.50	7.50	7.89	7.50
JSON 84KB	7.55	7.59	8.10	7.59
text 45KB	84.5	128	129	128
random 10KB	1.00	1.00	1.00	1.00
random 1MB	1.00	1.00	1.00	1.00
patterned 1MB	128	238	238	238

brotli, quality 6, lgwin 22: comprs vs node:zlib

Compression (MB/s of uncompressed data, higher is better)

Input	comprs	node:zlib
text 150B	15.4	12.4
JSON 84KB	77.6	76.6
text 45KB	1,576	1,462
random 10KB	108	101
random 1MB	423	415
patterned 1MB	1,335	1,328

Decompression (MB/s of uncompressed data, higher is better)

Input	comprs	node:zlib
text 150B	45.9	37.3
JSON 84KB	908	753
text 45KB	1,041	808
random 10KB	3,081	2,743
random 1MB	6,844	4,540
patterned 1MB	1,131	1,124

Compression ratio (input size ÷ output size, higher is better)

Input	comprs	node:zlib
text 150B	5.56	5.56
JSON 84KB	10.4	10.4
text 45KB	286	286
random 10KB	1.00	1.00
random 1MB	1.00	1.00
patterned 1MB	3,846	3,846

brotli, quality 11, lgwin 22: comprs vs node:zlib

Compression (MB/s of uncompressed data, higher is better)

Input	comprs	node:zlib
JSON 84KB	0.74	0.75
patterned 10KB	13.1	13.6

Decompression (MB/s of uncompressed data, higher is better)

Input	comprs	node:zlib
JSON 84KB	885	733
patterned 10KB	693	669

Compression ratio (input size ÷ output size, higher is better)

Input	comprs	node:zlib
JSON 84KB	13.5	13.5
patterned 10KB	46.1	46.1

zstd, level 3: comprs vs node:zlib

Compression (MB/s of uncompressed data, higher is better)

Input	comprs	node:zlib
text 150B	56.5	20.3
JSON 84KB	596	522
text 45KB	3,519	1,916
random 10KB	917	555
random 1MB	3,825	2,046
patterned 1MB	6,270	1,834

Decompression (MB/s of uncompressed data, higher is better)

Input	comprs	node:zlib
text 150B	103	36.5
JSON 84KB	1,757	1,280
text 45KB	7,917	2,952
random 10KB	4,753	2,124
random 1MB	14,178	4,506
patterned 1MB	17,091	4,801

Compression ratio (input size ÷ output size, higher is better)

Input	comprs	node:zlib
text 150B	4.84	4.84
JSON 84KB	9.25	9.25
text 45KB	285	285
random 10KB	1.00	1.00
random 1MB	1.00	1.00
patterned 1MB	2,755	2,755

Node.js streams: comprs vs node:zlib

Transform streams (MB/s of uncompressed data, higher is better)

Stream	Input	comprs	node:zlib
gzip compress, level 6	patterned 1MB in 16KB chunks	1,175	166
gzip decompress, level 6	patterned 1MB in 16KB chunks	1,710	329
gzip round-trip, level 6	patterned 1MB in 16KB chunks	762	108
zstd compress, level 3	patterned 1MB in 16KB chunks	1,646	388
zstd decompress, level 3	patterned 1MB in 16KB chunks	2,925	425
zstd round-trip, level 3	patterned 1MB in 16KB chunks	837	289

How they are measured

The Bench Report workflow runs scripts/bench-report.mjs on a GitHub-hosted runner, with every library at the same settings, and writes the tables and the charts from that one run, with the versions, the machine and the settings above the tables. Run the benchmarks yourself with pnpm run bench in a clone of the repository; numbers vary by machine and data. CodSpeed tracks the Rust benchmarks on the pull requests that change Rust code, to catch performance regressions.