Benchmarks

Faster than git clone. History on demand.

The Git source tree — 81,873 commits, 4,841 files — cloned from one laptop by three engines. Vex finishes in a median 10.40 s against git's 16.10 s from GitHub and 32.84 s from Entire, and produces a worktree verified against a git checkout of the same revision.

1.4x faster than git clone6x less data on the wireVerified against git checkout

Fine print: GitHub's own clone times swing widely with CDN and network conditions (3.13–12.36 s medians within a single day), so wall-clock comparisons here are same-window medians against the same repository and revision.

vex clone

Vex — shallow, deferred

Full commit history metadata plus the file content your working copy needs now. Any other file version streams in the moment you touch it.

git clone

git — full history

Git’s default: the complete history with every historical file version, whether or not you ever open one.

entire repo clone

Entire — full history

A git clone against an Entire cluster mirror, carrying the same complete history.

Bytes on the wire

Half the wait, a sixth of the data

The clone is quicker because Vex moves far less to do it: 51.7 MB against the 302 MiB git receives for the same tree. History and older file versions stream in on demand, the moment you actually open one.

Fast by architecture

Built to do less work, not just faster work

0.3 s

Commit index across 81,873 commits

A fresh clone bootstraps its index from the repository head instead of walking the entire operation history, so even deep histories index in a blink.

0 s decode

Streamed, chunk-framed packs

Pack chunks are independent zstd frames decoded the instant they arrive, so decompression and object writes hide entirely behind the download — and worktree files reflink from the local object store where the filesystem supports it.

On demand

Any file version, when you touch it

The default clone keeps full history metadata local, so logs and graphs are instant — and the content of any historical file version streams in the moment you open it.

Correctness first

No sample counts unless the worktree is byte-identical

Canonical digest

Every Vex and git worktree is hashed with the vex-normalized-tree-v1 canonical digest — raw file content, executable bits, and symlink targets. Every retained sample, default and eager, matched git's digest exactly.

Native, not emulated

Clones must report zero Git-compat fallback counters and at least one native trunk resolution. A sample that touches the Git compatibility path is invalid regardless of speed.

712 files, verified

Each sample materializes the complete 4,841-file worktree at the pinned revision — the shallow default included — and reproduces the same digest as a fresh GitHub clone of the same commit.

Methodology

How we measure

These are labelled production benchmark records, not synthetic microbenchmarks. Vex and GitHub run different services and transport layers, so speed ratios are honest within a day and indicative across days.

  • Three samples per engine from one MacBook Pro; medians published. The Vex figure was re-measured on 2026-08-08 after further clone work; the comparison engines are from the 2026-08-07 run, so the two were not sampled in the same window.
  • Fresh destination directory and client cache for every sample.
  • One fixture: the Git source tree at 2c78326f, 81,873 commits and 4,841 files — e2e-test-org/git on Vex, git/git on GitHub, /gh/PandelisZ/git on Entire.
  • Correctness gate: the Vex worktree is byte-identical to a git checkout of the same revision in 4,838 of 4,841 files. The three that differ are .bat files where git applies .gitattributes eol=crlf on checkout and Vex writes the raw blob; normalizing line endings makes all three identical.
  • Production infrastructure (vex.sc), August 2026. Wall-clock comparisons are only made within one measurement window; GitHub medians have swung several-fold across days in earlier records.

Clone it yourself

The fastest way to check a benchmark is to run the clone.

Get startedRead the docs