Speed

Fast, measured.

Speed is the point of VXO, so every number here comes from a benchmark that runs before every release. It has 36 budgets. If one is missed, the release stops.

36/ 36
Budgets met at the release gate, October 20, 2026.
247ms
Start the app to the first file rendered, p50 of 5.
11.7ms
Find a match 99% of the way through a 100 MB log.
33,774req/s
What the core answers with 16 clients at once.

Opening a file

The same probe for every app: ask macOS to open a file the way Finder does, and time it until the window is on screen.

Not running: double-click to window

Lower is better.

Cold open time in milliseconds
App and filems
VXO, a text file181
VXO, a PDF208
TextEdit, a text file222
Preview, a PDF371

Medians of four runs, scripts/winprobe.swift. Apple M3, macOS 15.3.2, October 20, 2026.

Already running: the next file

VXO opens a tab; TextEdit and Preview open a new window.

Warm open time in milliseconds, low and high
Applowhigh
VXO5.16.7
TextEdit2641
Preview87109

VXO: release gate p50s through the socket (5.1) and vxo FILE (6.7). Others: four runs. Apple M3, October 20, 2026.

Where a cold start goes

One launch, split by VXO's own timers (VXO_TIMINGS=1). Most of it is macOS starting its web engine. VXO's own work before the window is about 0.2 ms.

Cold start phases in milliseconds from launch
Phasefromto
macOS starts the app and its window062
Web view built62103
WebKit processes start103185
Page loads185195
File rendered195214

A typical launch from the low end of each phase's measured range, Apple M3, October 17, 2026 (SPEC section 11, item 16). Bench p50 of 5 at the release gate: 247 ms.

This is why VXO keeps running when you close its last window, the way Preview does: the second file of the day skips all of it and takes about seven milliseconds.

Always current

Save a file VXO is showing and time it until the new version is in the window. Every way an app can save, every release.

Saved to on screen

p50 of 5 each, in a real window. The budget is 50 ms.

Live reload in milliseconds
Kind of savems
A line appended (a growing log)12.6
Temporary file, then renamed (most Mac apps)15.4
Written in place20.7

Release gate, Apple M3, October 20, 2026. Thirteen recorded editor saves (vim, VS Code, NSDocument, sed, git checkout and more) are replayed in the tests; only the final version is ever shown.

Big files

Each bar is the measured time against its budget. A full bar would mean the budget was used up.

Time to the first screen

Cold, nothing cached by VXO.

Time against budget in milliseconds
Filemeasuredbudget
1 GB log, first page64.95300
200k-line Rust file, jump to line 190k92.70150
100 MB log, find near the end11.7150
1M-row xlsx, first page5.48300
100k-row CSV, first page with index6.58150
1 MB Markdown, rendered15.7930

Release gate, Apple M3, October 20, 2026.

Memory

Physical footprint, what Activity Monitor shows.

Memory against budget in megabytes
Statemeasuredbudget
Idle2.6930
With a 1 GB log open10.6360
App with a window open17.6330
With a 1M-row xlsx open42.5860
Peak of a 1 MB Markdown render6.3910

Medians after a settle period. Release gate, Apple M3, October 20, 2026.

Every budget

The full benchmark from the release gate. The bar under each number shows how much of its budget it used.

MetricMeasuredBudgetResult
Size
Release binary size4.35 MB≤ 8 MBPass
UI payload (index.html)91.81 KB≤ 96 KBPass
Opening files
Startup to UI served (p50)3.27 ms≤ 20 msPass
Window: spawn to first render (p50 of 5)246.68 ms≤ 400 msPass
Warm open: Finder route (LaunchServices) to content (p50 of 5)77.06 ms≤ 100 msPass
Warm open: vxo FILE to content in the window (p50 of 9)6.72 ms≤ 10 msPass
Warm open: socket hand-off to content in the window (p50 of 9)5.10 ms≤ 10 msPass
Live reload
Live reload, window: in-place save to on screen (p50 of 5)20.65 ms≤ 50 msPass
Live reload, window: temp file + rename to on screen (p50 of 5)15.35 ms≤ 50 msPass
Live reload, window: append a line to on screen (p50 of 5)12.64 ms≤ 50 msPass
Text, code and logs
100 MB log: first page (cold index, p50 of 5)8.62 ms≤ 250 msPass
100 MB log: 200 lines random seek (p95)0.24 ms≤ 5 msPass
100 MB log: find, hit 99% in or wrapped (p50 of 10)11.71 ms≤ 50 msPass
200k-line .rs: first highlighted page at line 190k (cold)92.70 ms≤ 150 msPass
200k-line .rs: 200 highlighted lines, random page (p95)0.50 ms≤ 3 msPass
1 GB log: first page (cold index, file in OS cache, p50 of 3)64.95 ms≤ 300 msPass
100 MB file: 1 MB Range read (p95)0.76 ms≤ 10 msPass
Spreadsheets
100k-row CSV: first page incl. index (cold, p50 of 5)6.58 ms≤ 150 msPass
100k-row CSV: deep page, warm index (p95)0.34 ms≤ 15 msPass
100k-row xlsx (6 styled cols): first page, cold (p50 of 5)1.87 ms≤ 300 msPass
100k-row xlsx: whole sheet parsed, cold (p50 of 5)146.17 ms≤ 400 msPass
100k-row xlsx: page, warm (p95)0.54 ms≤ 5 msPass
1M-row xlsx: first page, cold5.48 ms≤ 300 msPass
1M-row xlsx: whole sheet parsed1489.42 ms≤ 5000 msPass
1M-row xlsx: page anywhere, warm (p95)0.52 ms≤ 5 msPass
Markdown
1 MB markdown to HTML, cold (p50 of 5)15.79 ms≤ 30 msPass
1 MB markdown, cached (p95)1.17 ms≤ 5 msPass
Memory
Idle memory2.69 MB≤ 30 MBPass
Memory after 100 MB log8.17 MB≤ 60 MBPass
Memory after 1 GB log10.63 MB≤ 60 MBPass
Memory after load, settled 1 s (median of 5)11.55 MB≤ 60 MBPass
Memory with 100k-row xlsx open5.89 MB≤ 60 MBPass
Memory with 1M-row xlsx open42.58 MB≤ 60 MBPass
1 MB markdown: peak memory of a cold render6.39 MB≤ 10 MBPass
Window mode: vxo process memory (footprint)17.63 MB≤ 30 MBPass
Server
Throughput, 16 clients33,774 req/s≥ 2000 req/sPass

vxo 0.4.0, release gate reports/qa-20261020T095851Z.md, Apple M3, macOS 15.3.2, October 20, 2026. The latest development run (October 26, with other builds on the same Mac) met 34 of 36: warm open measured 16.3 ms and 14.0 ms against 10 ms. The gate is rerun on a quiet Mac before every release, and a miss holds it.

How we measure

The benchmark starts the real app from an app bundle, the way you run it. It times from the moment a command starts or a file is saved to the moment the window reports, with its own clock-stamped signal, that the new content is in it. Memory is what Activity Monitor shows. Nothing is simulated.

Some numbers have limits we can't move. The first window of the day waits for macOS to start its web engine, which takes about a tenth of a second on its own. That is why the first open takes a quarter of a second and the second takes seven milliseconds. Your Mac will differ; the method won't.

The benchmark is vxo-bench in the source, run by scripts/qa.sh together with the format, clippy, unit, end to end, UI and real-window tests.