Benchmarks

Benchmarks

Every number here is measured, reproducible, and gated on correctness — including the cells where a mojo kernel is not the fastest option. This page consolidates the per-package measured tables; each package README in the repo carries its own full table with per-cell latencies.

Methodology

• Correctness gate before timing. No timing pass runs until the kernel's output has been asserted against the real reference package at that kernel's documented tolerance (several are bit-identical). A fast wrong answer is a failed benchmark, not a fast one.
• Median of 5 runs for every cell (cold-start cells: median of 5 fresh processes, or ×3 patterns for search workloads). Absolute latencies shift with machine load; speedup ratios within each table are internally consistent.
• Cold and warm reported separately. Cold = first call in a fresh process (dlopen, runtime init, allocator warm-up). Warm = steady state. The consolidated table below cites warm numbers; notable cold regressions are flagged in the lowlights.
• Seeded, synthetic corpora. All inputs are generated locally from fixed seeds (Zipf-ish vocabularies, typo-seeded patterns, seeded MO coefficients, seeded soundings), so runs are reproducible and no upstream dataset is redistributed.
• Same machine, pinned environment. Unless a package README says otherwise: Apple M4 Max, macOS 26.6.2 arm64, Python 3.12.14, NumPy 2.5.3, Node v23.10.0, Mojo 1.1.0, measured 2026-09-19. Reference packages are pinned PyPI/npm releases (each README names its oracle version).
• Reproduce: pixi run bench, pixi run bench-cclib, pixi run bench-fuse for the three flagship kernels; benchmarks/bench_<name>.py / .mjs for the rest. The README tables are regenerated from these scripts, never hand-edited.

Consolidated measured table

Warm steady state, peak-and-range per package. "Parity" is the measured agreement against the reference oracle, asserted on both the native and forced-fallback backends.

Package	Accelerates	Measured speedup (warm)	Parity
bm25-mojo	rank_bm25	116×–8,769× (1k–100k docs)	max abs diff 3.6e-15
cclib-mojo	cclib method.volume	72× vs fastest shipping backend; 1,898–7,033× vs PyQuante1 path	max abs diff 1.5e-12
pypdf-filters-mojo	pypdf filters	809–1,776× LZW; 12–15× PNG/TIFF predictor end-to-end	identical decoded bytes
pykalman-mojo	pykalman	5–827× filter; 3.9–589× smooth	within 1e-10
elephant-mojo	elephant surrogates	218–342× dither; 2× p-value spectrum	bit-exact spectrum; dither ≤1e-3
uproot-mojo	uproot	172.6× vector<string> walk; 9.4× cold file→array	byte-for-byte
metpy-mojo	MetPy cape_cin	156.6× (625-col grid); 78× single column	within documented tolerances
natural-mojo	natural edit distances	3.7–64.5× (8–1,024 code units)	bit-identical distances
dynesty-mojo	dynesty	4.3–51.9× (defaults → bootstrap=0 vectorized)	posterior/evidence parity, asserted
jsonpath-mojo	jsonpath_ng	1.7–49.2× (shrinks with document size)	same matches, asserted
croniter-mojo	croniter	15.3–46.0× per get_next call	same fire times, asserted
obspy-mojo	ObsPy Konno–Ohmachi	3.9–6.7× steady-state; up to 45× with window build	rtol 1e-9, atol 1e-12 (float64)
nx-mojo	networkx	19.9–40× betweenness; 1.2–2.1× Dijkstra	asserted vs networkx, both backends
fuse-mojo	Fuse.js 7.1.0	8.8–38.7× warm; 2.9–13.4× cold	bit-identical (0 diff)
ruptures-mojo	ruptures	2.2–48× (Dynp 35–45×, PELT 10.5–48×, Binseg 2.2–2.9×)	same breakpoints, asserted
ase-mojo	ASE neighbor list	20.3–28.3× native (8.5–12.2× NumPy fallback)	identical (i, j, S) sets
jsonschema-mojo	jsonschema	2.1–3.1× small docs; 21–28× large nested	same verdicts and error surfaces
sacrebleu-mojo	sacrebleu	14.1–15.2× chrF; 2.9–3.1× BLEU end-to-end	0.0 measured diff (gate 1e-9)
toml-mojo	tomllib / tomlkit	1.8× vs tomllib; ~8× vs tomlkit (64 KB–1 MB)	same parsed objects, asserted
langdetect-mojo	langdetect	4.3–6.9× per detection	same detections, asserted
vader-mojo	vaderSentiment	4.1× warm	same scores, asserted
ckmeans-mojo	simple-statistics ckmeans	1.6–3.4×	exact cluster assignments
minisearch-mojo	MiniSearch fuzzy/prefix	0.8–3.1× by query mix	identical results, asserted
bio-mojo	Biopython Bio.SeqIO	1.4–2.3× (GenBank 2.1–2.3×)	record-for-record, asserted
ta-mojo	pandas-ta-classic	1.8–93× (indicator- and length-dependent)	bit-exact native; battery max 8.5e-14
jmespath-mojo	jmespath	0.01–1.1× — serialization-bound	exact incl. exception classes (5,000+ cells)

Honest lowlights

We publish the cells where the kernel loses, because the parity harness reports them whether we like the answer or not:

• jmespath-mojo is mostly ≤1× on its own benchmark. The kernel's raw evaluation is several times faster than the reference interpreter, but marshal.dumps serialization across the FFI boundary eats the win on typical queries (best cell: 1.10× on a 5-condition filter; worst: 0.01×). It ships because parity and the fallback are exact — not because it wins. This is what an honest 1.1× looks like.
• minisearch-mojo dips to 0.8–0.9× on exact-match and small fuzzy mixes at 10k–50k docs; it only earns its keep on prefix/autoSuggest mixes (up to 3.1× at 50k).
• Cold starts can cost more than they save on small work: ta-mojo's fastest indicator (EMA) measures 0.4× cold; vader-mojo is 0.85× cold (import + first call); uproot-mojo warm steady-state full open+read is 0.9× — its win is the cold path and the basket walk.
• bm25-mojo vs bm25s (numba incumbent): bm25-mojo v2 wins every cell of the 9-cell grid — 1.10×–1.95× warm steady-state (100k/20t re-verified at 1.12× over 11 interleaved samples) — plus 10–1000× on cold starts (AOT vs numba's ~4 s JIT). The v2 design: idf-weighted CSR baked at index time, query batching, no-rezero accumulation (float32 storage was measured and rejected: 94× over the 1e-8 parity gate, so float64 stays and is bit-faithful to rank_bm25 at 3.6e-15). Full autopsy and table in Kernel: bm25 and benchmarks/AUTOPSY-bm25s.md in the repo.
• Scaling caveats: pykalman-mojo's 827× is a small-system number (n_s=2); at n_s=20 the win is 5×. jsonpath-mojo slides from 49× to 1.7× as documents grow. sacrebleu-mojo's BLEU is bounded by the shared Python tokenizer (~3× end-to-end; the kernel's n-gram statistics alone are ~20×).

If a workload of yours shows a regression we didn't measure, that's a bug report we want: open an issue with the corpus shape.

---

Kernel catalog · How It Works · Apache-2.0, © 2026 Algenta