Benchmarks compare MoonBit encode/decode/load performance against Python tokenizers on the same corpora.

Charts

Performance Summary

Pipeline

Commands

python3 scripts/fetch_models.py
pip install tokenizers numpy
python3 scripts/gen_parity.py

moon bench --target native
python3 scripts/bench_compare.py --target native --corpus mixed
python3 scripts/bench_compare.py --target native --corpus all --fail-above 1.10

# Generate ECharts from benchmark report
node scripts/gen-bench-charts.mjs reports/bench-native-mixed.json

Reading Results

Moon/HF ratio	Interpretation
< 0.90x	MoonBit is faster on this case
0.90x .. 1.10x	Same range
> 1.10x	Optimization candidate or regression

Published performance claims should quote the comparison ratio, not standalone moon bench output.

The page reads /benchmarks/latest.json at runtime. CI writes the raw reports/bench-native-mixed.json artifact from bench_compare.py --json-out, then the docs build converts that report into the static JSON consumed here. ECharts are generated from the same report via gen-bench-charts.mjs.

Edit this page

Contributors: howtomakeaname

Next pageMethodology