Benchmarks

Apple M1 Pro, 2026-10-31.

5.9xfaster on Node 22 (short)

3.1-3.2xfaster on Node 24

3.0-3.3xfaster on Bun 1.4.4

1.48 MBdarwin-arm64 binary

thaiwrap vs Intl.Segmenter wrapper

Average time per call, lower is better. "short" is the 38-character example. "long" is the same text repeated 300 times. Bars are scaled to the reference.

Node 22.19.0 short

thaiwrap 3.35 µs

Intl.Segmenter 19.80 µs

5.91x

Node 22.19.0 long

thaiwrap 869 µs

Intl.Segmenter 5,010 µs

5.76x

Node 24.22.2 short

thaiwrap 3.46 µs

Intl.Segmenter 11.00 µs

3.18x

Node 24.22.2 long

thaiwrap 859 µs

Intl.Segmenter 2,630 µs

3.06x

Bun 1.4.4 short

thaiwrap 2.90 µs

Intl.Segmenter 9.70 µs

3.34x

Bun 1.4.4 long

thaiwrap 714 µs

Intl.Segmenter 2,140 µs

3.00x

The reference is the wrapper used in the tests. It breaks at word boundaries, so it does the same work as thaiwrap only on Thai-letter text such as these inputs.

Pure Rust core

Benchmark	Input	maxLength	µs per call
short	1x sample (38 chars)	10	1.89
long	300x sample	10	555.62
mixed	50x Thai/Latin/emoji	20	115.90
overflow_max4	300x sample	4	760.21
graphemes_long	300x sample, graphemes	10	632.32

criterion mean, median of 3 runs. Single machine, not guaranteed idle. Numbers from docs/m0/benchmarks.md.

Run it yourself

bun run build
bun run bench          # Bun
node bench/index.mjs   # Node
cargo bench -p thaiwrap-core