Real Workloads.
Real Results.
From data processing to machine learning, Pyvorin delivers measurable speedups across a wide range of Python workloads.
Measured · 2026-09-28 Geometric-mean speedup vs CPython 3.12.3 across 30 deterministic synthetic workloads in five categories (this host, 12 cores; per-workload median of 3 suite runs, each a median of 10 timed runs after 3 warmups; bar height ∝ √speedup). Every workload includes data generation in the timed region and must match CPython output exactly to count. Native-code coverage: 30/30 workloads correct and 0 fallbacks on the 2026-09-28 median-of-3 run (build 2026-09-28-v634). The 2026-09-18 caveat is superseded: the four ML rows it named are native again, and two native performance regressions it did not list (ds_summary_stats_1m 0.94x, agg_conditional_sum_1m 0.95x) are fixed to 18.1x and 15.7x. One row, web_log_tokenize_count, sits at 1.07x on a 1.03-1.17x distribution - the only row below 1.1x, and still faster than CPython. One ETL row (etl_csv_parse_conditional_sum) runs a native scan with a CPython-executed data-generation preamble; at publish the other 29 were fully native. The Mutable State bar is the dict write-back suite, measured 2026-09-09. Our homepage headline uses this same 30-workload suite (median of 3 runs, 2026-09-28). Raw results (JSON, run 1 of 3, with a provenance block naming the producing compiler build) · per-category summary (median of 3 runs) · mutable-state (dict write-back) suite.
Benchmarks you can trust.
Every published result ships with the workload source, dataset size, hardware specification, run count, median and p95, and a correctness check proving Pyvorin output matches CPython exactly. No cherry-picking, no mystery numbers.
| Workload | Category | Type | Status | Speedup |
|---|---|---|---|---|
| 30-workload morphic suite (6 per category) | All categories | Synthetic · deterministic | Measured 2026-09-28 | 5.00x geomean (1.07x-395x) · measured 2026-09-28 |
| ETL / data processing | Data Engineering | Synthetic | Measured 2026-09-28 | 4.4x geomean · fully native (1 hybrid row) |
| ML kernels (k-means, softmax, regression…) | Machine Learning | Synthetic | Measured 2026-09-28 | 4.6x geomean · fully native |
| Numeric analytics (stats, rolling, percentiles) | Data Science | Synthetic | Measured 2026-09-28 | 9.8x geomean · fully native |
| String/dict-heavy service patterns | Web Services | Synthetic | Measured 2026-09-28 | 2.1x geomean · native (1 hybrid row) |
| Aggregations (sums, group-totals, multi-reduce) | Aggregations | Synthetic | Measured 2026-09-28 | 7.7x geomean · fully native |
| Stateful dict write-back (scoring, ledger posting, revaluation, session decay, event rollup, RMW balance adjustment) | Mutable State | Synthetic | Measured 2026-09-09 | 1.8x geomean · 100% native at publish · 1.5x additional lift from the fused dict bridge |
Internal 50-workload suite — two rows published on the CPython fallback
The published dataset now covers the full 50-workload registry (it previously held 48). The two rows below were added on 2026-09-18 and are not native: both execute on the CPython fallback path on the current build, which is why they sit at ~1.0x. They are published explicitly labelled rather than dropped. No re-measurement was made on 2026-09-18 — this host was contended, so a fresh ratio would not be evidence — so the figures are the values measured 2026-09-17. Compiler build stamped at publish: 2026-09-17-wave5-v586.
| Workload | Category | Type | Status | Speedup |
|---|---|---|---|---|
| nbody_200k | Scientific | Synthetic · deterministic | CPython fallback · not native | 0.99x · not native (CPython fallback, MP-83) |
| spectral_norm_500 | Scientific | Synthetic · deterministic | CPython fallback · not native | 1.02x · not native (CPython fallback, MP-82) |
Don’t take our word for it. Measure it.
The trial includes pyvorin bench - run it against your own code and see your own numbers.