Benchmark work follows three rules:

1. Compare MoonBit and Python tokenizers on the same model and corpus.
2. Keep correctness fixtures passing before interpreting speed.
3. Treat Moon/HF > 1.10x as a concrete optimization backlog item.

Corpora

Corpus	Purpose
short	Interactive overhead and chat-style snippets
mixed	English, CJK and punctuation paragraph
code	MoonBit-like code and identifiers
long	Long-document BPE/Unigram stress

Outputs

bench_compare.py can write JSON artifacts for CI or dashboards:

python3 scripts/bench_compare.py --target native --corpus all \
  --fail-above 1.10 --json-out reports/bench-native-all.json

For the documentation site, convert a raw benchmark report into the static snapshot consumed by the performance page:

node scripts/update-docs-benchmarks.mjs reports/bench-native-all.json

CI keeps the raw reports/bench-native-mixed.json as an artifact, while Pages uses docs/.vuepress/public/benchmarks/latest.json for the rendered chart and table.

Edit this page

Contributors: howtomakeaname

Previous pageBenchmarks