Benchmark

How we tested,
and every result.

The comparison on the technology page is a summary. This page has the setup, how each number was measured, all the results (including the ones where Leyr is not the best), the files to download and the steps to run it yourself.

Who ran it

The Leyr project, so this is not independent

We wrote the harness, chose the library and the settings, and ran the tests. That is a reason to check our work. Everything needed to rerun it is on this page. If you rerun it and get different numbers, send them to us and we will publish them next to ours with your setup.

Setup

Machine, software and limits

The machine is a desktop that also ran other work during the tests, so timings carry noise (see the caveats).

Run by
The Leyr project. This is not an independent test.
Dates
2026-11-14, 2026-11-15
Hardware
AMD Ryzen 9 5950X 16-Core Processor, 32 threads, 47 GB RAM; music on local NVMe
Operating system
Arch Linux, kernel 7.3.4-arch1-2
Container engine
podman version 6.2.0, rootless, crun runtime
Limits
Every container: 4 CPU cores, 4 GB of memory, no swap (--cpus=4 --memory=4g --memory-swap=4g), one container at a time.
Network
Every container runs on an internal podman network with no route out (podman network create --internal). Published ports still reach the host. Before each run the harness tries to reach three outside addresses from inside the container and stops if any answers; each result file records that check.
Leyr
main 0504a876 plus the memory change (commit a105ba12; image built from the repository Dockerfile, image id c21eb7c848fb)
Rerun
On 2026-11-15 (evening) Leyr and gonic were rerun together on the same library and cores, with a build of Leyr from main before the change and one after it taken in between (the before-change runs are in results/memory-333-before). Navidrome, Jellyfin and Plex were not rerun: their rows are from earlier runs on the same library, and Jellyfin and Plex have private memory at idle only. The host was busy (load average 12 to 24) during the whole rerun, so timings are noisy. The newer Leyr build changes memory only: its request and stream timings match the earlier build within run-to-run noise.
CPU pinning
Rootless podman has no cpuset controller, so every container's processes were pinned to four fixed cores with taskset after start, on top of the 4-CPU limit, away from the cores another job on the machine was pinned to.
Other load
The machine was shared. The 1-minute load average before each run is in each library's progress.log inside the results download; it ranged from about 5 to 63. Other software (compiles and tests by other jobs on the same machine) ran during the runs.
Clients
The harness runs on the same host, in Python, over HTTP on 127.0.0.1.
Server images, pinned by digest.
ServerVersionImage
Leyrbuilt from the commit belowlocalhost/bench-leyr
Navidrome0.64.2 (10114574)docker.io/deluan/navidrome@sha256:38dc2727bfcfd5ede290f8ada114fc90368146f265ae4701ddddbcbe2a44ee52
gonicv0.23.0docker.io/sentriz/gonic@sha256:7b2298d905149ecb7e57cdd3bf7a40082758b9bda2e24454b41c17c99bf25e8f
JellyfinJellyfin.Server 12.4.0.0docker.io/jellyfin/jellyfin@sha256:fea863d0cd9ec445985d9c75fc6ed64e38a1d2766cf0c250a33d553f4fc1b771
Plex1.43.5.11079-b3b13e506docker.io/plexinc/pms-docker@sha256:6d778d41daae68f89115aa87a44ae14a96b23e71a2093da173b5a9aac4091336

Settings

What differs from the defaults, and why

The aim was the same job for every server and nothing leaving the machine. Each of these changes could favour or hurt a server, so they are listed.

Settings changed from each server's defaults.
ServerChangeWhy
AllInternal network, no route out; 4 CPUs, 4 GB; music mounted read-only; fresh volume per scan runThe same conditions for every server, and private music cannot leave the machine.
LeyrSOULSEERR_BACKGROUND_JOBS=off, LEYR_REQUEST_QUEUE=off, LEYR_ROOMS_RUNTIME=off, LEYR_EXPORT_RUNTIME=off, LEYR_BACKGROUND_RUNTIME=off, LIBRARY_SCAN_MINUTES=0Turns off jobs that call outside services (downloads, Discord, MusicBrainz) or run on a timer. The first scan runs at boot regardless.
NavidromeND_SCANNER_SCHEDULE=0, ND_ENABLEEXTERNALSERVICES=false, ND_LASTFM_ENABLED=false, ND_ENABLEINSIGHTSCOLLECTOR=falseNo timed scans and no outside lookups or telemetry. The scan is started through its API.
gonicScan interval, file watcher and scan-at-start off; playlist and podcast folders on volumesNo timed scans; gonic refuses to start without those folders.
JellyfinRealtime monitor off, internet providers off, and the metadata and image fetchers for artists, albums and tracks emptiedNo outside lookups. Side effect: Jellyfin then extracts no album art, so its cover-art rows are n/a. With the providers on, the scan took 1,066 s when they could not connect.
PlexUnclaimed (no account); ALLOWED_NETWORKS=0.0.0.0/0 so the API answers without a token; the library uses the Personal Media agent; loudness and sonic analysis, thumbnail generation, library auto-scan, DLNA and server publishing offNo account is needed or created. The Plex Music agent looks albums up on Plex's servers; the Personal Media agent reads tags and embedded art only. Plex also downloads its MP3 encoder from plex.tv the first time it converts a track, which cannot happen with no network, so every Plex conversion row (MP3, Opus, seek) is n/a. We did not open the network for it, to keep to the rule that nothing leaves the machine.

Measurements

What was measured, and how

One harness (bench.py, Python standard library only) drives each server in turn. Every server gets two full runs and two extra scan-only runs on each library.

Start-up
Time from podman start to the first good answer, in milliseconds: Leyr /api/health, Navidrome /ping, gonic a Subsonic reply, Jellyfin /Users/Public (its /health answers before it is ready), Plex /identity once it stops reporting startState. First boot is an empty volume. Restarts use the populated database, five per full run.
Scan
Fresh volume, create the admin (or, for Plex, the library), start a full scan, and poll until the server says it is finished, the track count has stopped changing and the container used under 0.15 CPU-seconds in a 6-second window. The time ends at the start of that quiet window. This matters most for Jellyfin, which keeps working after its scan task ends. Plex starts scanning when its library is created, which the harness does as the scan step. Leyr starts scanning on its own at boot and finishes in a second or two, before the harness can poll, so for Leyr the scan time is the time between two lines of Leyr's own log ("native music library open" and "library scan finished"), and the harness still waits for the container to go quiet. For the others the time includes up to a second of polling. What each server does during a scan (art, analysis, search indexes) is not the same work.
Memory and threads
Resident memory (RSS) read from /proc, added up over every process in the container, sampled every 0.5 s during the scan and every 0.25 s under load. Idle is one reading 300 s after the scan ends. Threads are the same sum. Private memory is the Anonymous total of each process's smaps_rollup, summed the same way and sampled at the same moments: heap, stacks and other pages the process owns that no file backs. RSS also counts the executable, libraries and mapped database files; the kernel can take those back whenever it needs memory, and how many are mapped at a given moment varies from run to run, so private memory is the steadier measure of what a server holds. The earlier Jellyfin and Plex runs recorded it at idle only.
API latency
Six calls per server from the host over keep-alive HTTP, after 40 warm-up requests each: 200 requests one at a time, then 800 across 16 connections. Requests rotate through 20 albums and 10 search terms. Median and 95th percentile in milliseconds. The Python client adds a little time of its own, the same for every server, and tops out near 2,000 to 4,000 requests a second.
Streams
Time from sending the request to the first audio byte, on a new connection each time: the original FLAC; FLAC converted to MP3 at 320 kbit/s and to Opus at 128 kbit/s (20 different tracks each, so a cache cannot answer a repeat); and on the 63-minute FLAC the original, MP3 from the start, and MP3 and Opus starting at 25:00 (a different offset for each request, 1500 s plus 7 s per request). Repeating the same request ten times is reported separately, because Navidrome and Jellyfin answer repeats from a cache.
Seek check
validate_seek.py downloads the start of the converted stream at offset 0, 1500 s and 1507 s and decodes two seconds of each, comparing the loudness with the source file at the same positions. A server that ignores the offset fails; Jellyfin did, with the request form used, so its seek rows are n/a. Plex could not be checked: with no network it cannot convert audio, so its line in the seek file is empty (?).
Memory under load
The same sampler during the 16-connection runs of all six calls.
The request behind each row. Jellyfin and Plex have no Subsonic API, so they get their closest native request.
RowLeyr, Navidrome, gonic (Subsonic)JellyfinPlex
pingpingGET /System/Info/PublicGET /identity
getAlbumList2_newest50getAlbumList2 type=newest size=50GET /Items (albums, newest first, 50)GET /library/sections/1/all?type=9&sort=addedAt:desc (50)
getArtistsgetArtistsGET /ArtistsGET /library/sections/1/all?type=8
getAlbumgetAlbumGET /Items?ParentId=<album>GET /library/metadata/<album>/children
search3search3 (20 songs, 20 albums, 20 artists)GET /Search/Hints?limit=60GET /hubs/search?limit=20
getCoverArt_300getCoverArt size=300GET /Items/<id>/Images/Primary?maxWidth=300GET /photo/:/transcode?width=300&height=300
flac_originalstream format=raw/Audio/<id>/stream?static=truethe track's file URL (direct play)
flac_to_mp3_320, seek rowsstream format=mp3 maxBitRate=320 [timeOffset=1500]/Audio/<id>/stream.mp3?audioBitRate=320000 [startTimeTicks] (ignored)/music/:/transcode/universal/start.mp3?musicBitrate=320 [offset=1500]
flac_to_opus_128stream format=opus maxBitRate=128stream.opusn/a: no conversion without network

Libraries

Two libraries, the same tests

Every server ran on both libraries. The public one is the headline because anyone can rebuild it. The real one checks that the public results hold on real music.

Public library (generated)

374 albums, 3,098 tracks, 29.6 GB: 498 FLAC, 20 M4A, 2,580 MP3. Seed 20261113; generated with ffmpeg version n9.0.2 Copyright (c) 2000-2026 the FFmpeg developers in 17 minutes.

Made by make_library.py from a fixed seed and ffmpeg alone: sine tones and seeded noise, invented artists and album names, generated cover art. No recordings, nothing copyrighted. Tracks are 3 to 6 minutes, with tags (artist, album artist, album, title, track, disc, genre, year), compilations and featured artists, and an embedded cover in every file. One 63-minute FLAC ("Long Form Drone") is used for the seek tests. Checksums for every file are in the downloads.

Real library (private)

374 album folders from a private collection, copied to local disk: 3,129 audio files, 30.7 GB (2,587 MP3, 522 FLAC, 20 M4A) with embedded art, plus a 63-minute FLAC for the seek tests. It cannot be published, so its results cannot be checked by others. Only counts and the format mix are given here. build_subset.sh shows how a subset of any collection is copied.

Results

Public library, all results

Every row we measured, for every server. Where two numbers appear they are run 1 and run 2. The lowest value in a row is highlighted; a highlight inside the noise (about two times) is not a win.

Measured: Leyr 2026-11-15; Navidrome 2026-11-15; gonic 2026-11-15; Jellyfin 2026-11-15; Plex 2026-11-15. Leyr can be rerun alone, so its date can differ from the others.

Scan of the library

Seconds from trigger to finished and quiet. Every scan run is listed (the full runs count as scan runs).
WhatLeyrgonicNavidromeJellyfinPlex
Scan time (s)1.4, 1.0, 1.0, 1.0 (median 1.0)6.8, 7.6, 6.3, 7.5 (median 7.1)14.4, 10.8, 10.4, 10.4 (median 10.6)283.1, 290.1, 259.1, 264.4 (median 273.8)109.3, 145.1, 103.0, 98.1 (median 106.1)
Peak memory during scan (MB)41, 42, 43, 4350, 47, 55, 66168, 142, 153, 175502, 500, 507, 474493, 491, 494, 494
Private memory, peak during scan (MB)19, 20, 21, 2127, 25, 33, 45137, 114, 123, 145n/an/a
Tracks indexed30983098309830983098

Memory and threads

Resident memory (RSS) summed over every process in the container. Private memory is the Anonymous total of each process's smaps_rollup, summed the same way: heap, stacks and every other page the process owns that is not backed by a file, so it leaves out the executable, libraries and mapped database files, which the kernel can take back whenever it needs memory. n/a: the run was made before this was recorded. Run 1 ; run 2.
WhatLeyrgonicNavidromeJellyfinPlex
Idle memory, 5 min after the scan (MB)26 ; 3238 ; 4355 ; 56402 ; 361401 ; 362
Idle private memory (MB)14 ; 1523 ; 2024 ; 26230 ; 210245 ; 232
Idle threads4 ; 411 ; 1229 ; 2920 ; 19114 ; 111
Memory under load, median (MB)50 ; 5056 ; 6075 ; 74468 ; 434442 ; 436
Memory under load, peak (MB)58 ; 5663 ; 64120 ; 119483 ; 448453 ; 451
Private memory under load, peak (MB)30 ; 2542 ; 4188 ; 89n/an/a

Start-up

Container start to first good answer. First boot is an empty volume; restarts use the populated database (5 restarts per run).
WhatLeyrgonicNavidromeJellyfinPlex
First boot, empty volume (s)0.4 ; 0.40.3 ; 0.31.1 ; 1.14.3 ; 3.928.3 ; 32.1
Restart, median of 5 (ms)345 ; 313283 ; 310309 ; 2882,282 ; 2,9041,594 ; 1,861
Restart, slowest of 5 (ms)375 ; 524377 ; 356373 ; 2982,964 ; 3,6151,808 ; 2,098

API latency, one client at a time

Median/p95 in ms, run 1 ; run 2. 200 requests (one client) or 800 (16 clients) per run. Jellyfin and Plex use their own APIs, so their rows compare similar work, not identical calls.
WhatLeyrgonicNavidromeJellyfinPlex
ping0.3/0.4 ; 0.3/0.50.3/0.3 ; 0.3/0.40.5/0.6 ; 0.4/0.53.0/4.3 ; 4.2/5.90.2/0.4 ; 0.2/0.4
getAlbumList2_newest501.7/3.4 ; 1.2/1.65.4/7.6 ; 6.0/15.43.8/4.9 ; 2.7/3.617.0/20.2 ; 28.0/40.99.9/12.2 ; 11.0/15.3
getArtists1.9/3.1 ; 1.2/1.66.6/8.9 ; 6.8/16.05.2/6.2 ; 3.4/4.721.6/26.1 ; 45.4/70.85.8/8.7 ; 5.7/7.6
getAlbum1.1/2.4 ; 0.9/1.22.6/3.9 ; 2.5/6.52.1/2.8 ; 1.6/2.213.0/15.4 ; 18.0/23.03.3/4.3 ; 3.6/5.7
search35.1/8.3 ; 4.1/6.87.5/10.9 ; 6.9/15.97.6/9.0 ; 4.9/6.421.1/26.7 ; 28.9/38.132.0/43.6 ; 35.0/45.6
getCoverArt_3000.7/1.1 ; 0.6/0.70.4/0.6 ; 0.3/1.00.8/1.0 ; 0.6/0.7n/a ; n/a0.4/0.7 ; 0.3/0.6

API latency, 16 clients at once

Median/p95 in ms, run 1 ; run 2. 200 requests (one client) or 800 (16 clients) per run. Jellyfin and Plex use their own APIs, so their rows compare similar work, not identical calls.
WhatLeyrgonicNavidromeJellyfinPlex
ping1.9/5.1 ; 1.7/4.82.1/6.5 ; 2.5/9.55.0/22.3 ; 2.6/10.012.9/20.5 ; 17.1/27.71.8/5.8 ; 2.0/6.1
getAlbumList2_newest5011.2/17.2 ; 10.3/15.532.9/49.2 ; 40.2/67.154.5/128.5 ; 36.5/83.377.9/121.7 ; 137.1/224.0217.1/252.5 ; 235.4/268.7
getArtists11.8/18.1 ; 9.5/12.829.2/51.8 ; 37.8/71.457.9/118.4 ; 40.6/82.8116.7/185.6 ; 335.3/558.4116.8/142.7 ; 141.2/181.2
getAlbum8.8/12.6 ; 5.3/10.013.6/19.8 ; 15.3/24.340.1/109.6 ; 31.7/76.068.6/104.8 ; 164.7/275.477.6/106.0 ; 86.4/122.3
search338.3/57.8 ; 26.4/33.540.8/59.5 ; 43.3/58.699.2/252.0 ; 43.2/102.5110.5/169.7 ; 229.5/329.5530.8/4,658.9 ; 865.4/6,892.8
getCoverArt_3002.3/4.9 ; 2.5/5.82.6/5.6 ; 2.0/6.46.6/26.5 ; 4.5/21.7n/a ; n/a2.6/5.8 ; 3.0/6.4

Streams: time to first audio byte

Median/p95 in ms per series, run 1 ; run 2. 20 different tracks per series, 10 requests on the long file.
WhatLeyrgonicNavidromeJellyfinPlex
flac_original1.5/16.3 ; 1.1/15.20.9/1.5 ; 1.3/16.72.3/17.3 ; 1.8/16.611.5/76.2 ; 21.2/101.61.2/2.0 ; 2.3/6.3
flac_to_mp3_32059.9/209.6 ; 57.7/204.432.9/113.2 ; 31.2/113.2143.3/293.0 ; 105.2/212.5301.5/667.4 ; 421.8/707.4n/a ; n/a
flac_to_opus_12860.4/66.2 ; 57.1/124.448.0/54.3 ; 44.4/48.9718.0/948.8 ; 474.9/757.8732.8/1,566.2 ; 833.8/1,775.4n/a ; n/a
long_flac_original1.3/3.0 ; 0.9/15.30.8/1.8 ; 0.7/15.625.2/39.7 ; 13.9/53.423.0/48.0 ; 27.4/52.31.5/2.4 ; 1.7/4.4
long_flac_mp3_320_from_start_repeated57.8/65.3 (first 65) ; 59.5/72.1 (first 57)31.3/34.6 (first 30) ; 29.2/32.0 (first 29)27.1/543.9 (first 544) ; 28.1/521.9 (first 522)16.7/324.6 (first 325) ; 23.3/334.3 (first 334)n/a ; n/a
long_flac_mp3_320_seek_1500s_plus61.4/77.9 ; 55.6/75.0734.1/803.1 ; 889.6/1,027.2722.5/859.3 ; 570.1/649.7n/a ; n/an/a ; n/a
long_flac_opus_128_seek_1500s_plus58.5/63.8 ; 55.9/59.8951.0/1,253.9 ; 1,192.6/1,332.21,252.0/1,469.6 ; 914.6/1,031.1n/a ; n/an/a ; n/a
long_flac_mp3_320_seek_1500s_repeated59.7/66.8 (first 60) ; 53.5/57.1 (first 53)877.7/1,181.3 (first 818) ; 1,248.8/1,696.3 (first 1,344)11.5/43.4 (first 41) ; 4.8/40.0 (first 4)n/a ; n/an/a ; n/a

Where Leyr is not the best

Rows where another server had the lower median, or where Leyr was within 25% of the lowest (a tie, given the noise). Values are the median of the runs, in the row's unit.

  • Peak memory during scan (MB) (scan of the library): Leyr 42.8, gonic 52.7. Leyr is within 25% of the best.
  • Memory under load, median (MB) (memory and threads): Leyr 50.0, gonic 57.9. Leyr is within 25% of the best.
  • Memory under load, peak (MB) (memory and threads): Leyr 56.9, gonic 63.6. Leyr is within 25% of the best.
  • First boot, empty volume (s) (start-up): Leyr 0.4, gonic 0.3. Leyr is within 25% of the best.
  • Restart, median of 5 (ms) (start-up): Leyr 329.3, gonic 296.3. Leyr is within 25% of the best.
  • Restart, slowest of 5 (ms) (start-up): Leyr 449.7, Navidrome 335.6. Leyr is behind the best.
  • ping (api latency, one client at a time): Leyr 0.3, Plex 0.2. Leyr is behind the best.
  • getCoverArt_300 (api latency, one client at a time): Leyr 0.6, Plex 0.3. Leyr is behind the best.
  • ping (api latency, 16 clients at once): Leyr 1.8, Plex 1.9. Leyr is within 25% of the best.
  • getCoverArt_300 (api latency, 16 clients at once): Leyr 2.4, gonic 2.3. Leyr is within 25% of the best.
  • flac_original (streams: time to first audio byte): Leyr 1.3, gonic 1.1. Leyr is within 25% of the best.
  • flac_to_mp3_320 (streams: time to first audio byte): Leyr 58.8, gonic 32.0. Leyr is behind the best.
  • flac_to_opus_128 (streams: time to first audio byte): Leyr 58.7, gonic 46.2. Leyr is behind the best.
  • long_flac_original (streams: time to first audio byte): Leyr 1.1, gonic 0.8. Leyr is behind the best.
  • long_flac_mp3_320_from_start_repeated (streams: time to first audio byte): Leyr 58.6, Jellyfin 20.0. Leyr is behind the best.
  • long_flac_mp3_320_seek_1500s_repeated (streams: time to first audio byte): Leyr 56.6, Navidrome 8.1. Leyr is behind the best.

Results

Real library, all results

The same tests on the private library. These confirm whether the public results hold on real music; they cannot be reproduced by others.

Measured: Leyr 2026-11-15; Navidrome 2026-11-15; gonic 2026-11-15; Jellyfin 2026-11-15; Plex 2026-11-15. Leyr can be rerun alone, so its date can differ from the others.

Scan of the library

Seconds from trigger to finished and quiet. Every scan run is listed (the full runs count as scan runs).
WhatLeyrgonicNavidromeJellyfinPlex
Scan time (s)3.5, 3.4, 3.0, 4.6 (median 3.5)10.1, 24.0, 9.9, 18.0 (median 14.0)25.5, 30.0, 26.3, 30.3 (median 28.1)177.9, 251.9, 198.7, 237.5 (median 218.1)183.3, 245.7, 205.5, 191.2 (median 198.4)
Peak memory during scan (MB)72, 82, 88, 5971, 73, 72, 861,148, 1,131, 1,163, 1,122611, 607, 505, 527566, 553, 556, 557
Private memory, peak during scan (MB)n/an/an/an/an/a
Tracks indexed31223129312231223122

Memory and threads

Resident memory (RSS) summed over every process in the container. Private memory is the Anonymous total of each process's smaps_rollup, summed the same way: heap, stacks and every other page the process owns that is not backed by a file, so it leaves out the executable, libraries and mapped database files, which the kernel can take back whenever it needs memory. n/a: the run was made before this was recorded. Run 1 ; run 2.
WhatLeyrgonicNavidromeJellyfinPlex
Idle memory, 5 min after the scan (MB)59 ; 4555 ; 3464 ; 63486 ; 390366 ; 338
Idle private memory (MB)39 ; 2533 ; 2433 ; 32246 ; 220233 ; 232
Idle threads6 ; 612 ; 1229 ; 2618 ; 18113 ; 114
Memory under load, median (MB)77 ; 7371 ; 5588 ; 84551 ; 443414 ; 297
Memory under load, peak (MB)86 ; 8277 ; 60107 ; 115564 ; 464415 ; 374
Private memory under load, peak (MB)n/an/an/an/an/a

Start-up

Container start to first good answer. First boot is an empty volume; restarts use the populated database (5 restarts per run).
WhatLeyrgonicNavidromeJellyfinPlex
First boot, empty volume (s)0.4 ; 0.50.3 ; 0.31.0 ; 1.43.8 ; 5.146.9 ; 51.3
Restart, median of 5 (ms)300 ; 380313 ; 408294 ; 3862,225 ; 2,4051,525 ; 1,805
Restart, slowest of 5 (ms)309 ; 481350 ; 4363,316 ; 4362,285 ; 8,7082,524 ; 2,894

API latency, one client at a time

Median/p95 in ms, run 1 ; run 2. 200 requests (one client) or 800 (16 clients) per run. Jellyfin and Plex use their own APIs, so their rows compare similar work, not identical calls.
WhatLeyrgonicNavidromeJellyfinPlex
ping0.3/0.3 ; 0.3/0.50.2/0.3 ; 0.3/0.50.4/0.6 ; 0.7/1.22.8/3.8 ; 3.4/5.50.2/0.4 ; 0.3/0.5
getAlbumList2_newest501.4/1.9 ; 2.2/4.34.7/6.5 ; 7.0/10.52.9/4.7 ; 5.2/13.315.8/18.6 ; 17.5/20.48.6/10.4 ; 11.2/15.1
getArtists1.8/2.3 ; 3.2/5.47.5/8.9 ; 20.3/47.15.7/7.7 ; 9.8/22.420.5/24.1 ; 22.3/25.18.0/9.2 ; 11.1/14.5
getAlbum1.0/1.8 ; 1.5/3.42.4/3.5 ; 4.6/7.51.6/2.5 ; 3.7/10.713.4/19.3 ; 14.7/21.93.1/4.3 ; 3.7/6.2
search34.4/7.7 ; 6.3/10.65.4/7.7 ; 9.6/15.34.2/7.5 ; 8.3/21.020.3/23.5 ; 24.8/32.842.5/58.3 ; 53.5/75.2
getCoverArt_3000.5/0.6 ; 0.7/1.00.3/0.5 ; 0.5/1.10.6/1.0 ; 1.1/3.8n/a ; n/a0.3/0.3 ; 0.4/0.9

API latency, 16 clients at once

Median/p95 in ms, run 1 ; run 2. 200 requests (one client) or 800 (16 clients) per run. Jellyfin and Plex use their own APIs, so their rows compare similar work, not identical calls.
WhatLeyrgonicNavidromeJellyfinPlex
ping1.7/4.8 ; 2.3/6.31.8/6.1 ; 2.5/7.43.9/14.5 ; 5.6/23.511.7/17.3 ; 18.3/29.51.7/5.5 ; 2.0/6.3
getAlbumList2_newest5011.7/14.6 ; 18.7/28.826.9/40.5 ; 49.7/76.136.5/96.5 ; 106.9/319.379.9/114.8 ; 114.0/180.5160.4/188.0 ; 188.5/221.9
getArtists14.2/19.7 ; 16.5/24.738.5/66.7 ; 75.4/129.250.8/104.4 ; 236.8/417.2107.6/190.2 ; 163.1/271.9178.4/219.0 ; 236.7/287.8
getAlbum6.9/10.1 ; 8.4/13.812.7/16.7 ; 24.5/38.731.6/89.2 ; 129.1/301.566.6/105.1 ; 77.4/125.756.0/85.5 ; 78.9/125.7
search325.0/34.7 ; 32.9/42.832.5/49.0 ; 71.0/194.997.5/202.3 ; 91.9/220.299.7/149.5 ; 140.9/226.73,252.6/4,927.5 ; 4,109.9/6,611.5
getCoverArt_3002.2/4.6 ; 2.8/5.22.4/6.4 ; 4.7/11.08.1/28.4 ; 30.7/95.4n/a ; n/a2.9/6.2 ; 2.9/6.0

Streams: time to first audio byte

Median/p95 in ms per series, run 1 ; run 2. 20 different tracks per series, 10 requests on the long file.
WhatLeyrgonicNavidromeJellyfinPlex
flac_original0.8/6.4 ; 0.9/14.30.8/1.4 ; 3.7/17.31.7/2.3 ; 4.3/34.88.5/57.9 ; 10.7/56.01.2/16.2 ; 2.9/21.6
flac_to_mp3_32055.7/110.8 ; 70.5/254.633.3/115.4 ; 61.0/220.4124.2/217.1 ; 211.2/447.8399.6/982.9 ; 679.6/1,123.1n/a ; n/a
flac_to_opus_12857.6/75.6 ; 66.5/72.344.8/50.0 ; 78.2/90.2472.8/683.2 ; 705.6/954.3599.7/2,050.5 ; 777.7/3,182.8n/a ; n/a
long_flac_original1.0/2.8 ; 1.6/15.61.0/2.0 ; 1.7/6.68.6/39.4 ; 15.8/73.426.0/37.1 ; 58.6/95.21.1/2.3 ; 1.1/3.5
long_flac_mp3_320_from_start_repeated56.8/60.1 (first 58) ; 61.6/66.4 (first 58)29.7/33.1 (first 28) ; 53.5/61.4 (first 54)18.1/437.7 (first 438) ; 24.2/587.3 (first 587)21.6/430.5 (first 430) ; 62.6/409.1 (first 409)n/a ; n/a
long_flac_mp3_320_seek_1500s_plus58.9/83.1 ; 63.0/79.5647.0/700.2 ; 1,050.1/1,372.3459.0/602.6 ; 589.1/930.8n/a ; n/an/a ; n/a
long_flac_opus_128_seek_1500s_plus56.7/58.4 ; 60.7/66.3863.5/1,034.3 ; 1,356.3/1,711.0725.3/821.0 ; 807.9/883.8n/a ; n/an/a ; n/a
long_flac_mp3_320_seek_1500s_repeated54.7/56.1 (first 55) ; 62.9/66.7 (first 63)617.1/771.3 (first 686) ; 908.0/1,199.5 (first 1,027)9.7/38.5 (first 12) ; 7.7/56.7 (first 4)n/a ; n/an/a ; n/a

Where Leyr is not the best

Rows where another server had the lower median, or where Leyr was within 25% of the lowest (a tie, given the noise). Values are the median of the runs, in the row's unit.

  • Peak memory during scan (MB) (scan of the library): Leyr 76.8, gonic 72.5. Leyr is within 25% of the best.
  • Idle memory, 5 min after the scan (MB) (memory and threads): Leyr 52.1, gonic 44.7. Leyr is within 25% of the best.
  • Idle private memory (MB) (memory and threads): Leyr 31.8, gonic 28.0. Leyr is within 25% of the best.
  • Memory under load, median (MB) (memory and threads): Leyr 75.2, gonic 63.1. Leyr is within 25% of the best.
  • Memory under load, peak (MB) (memory and threads): Leyr 84.1, gonic 68.5. Leyr is within 25% of the best.
  • First boot, empty volume (s) (start-up): Leyr 0.5, gonic 0.3. Leyr is behind the best.
  • Restart, median of 5 (ms) (start-up): Leyr 339.8, Navidrome 340.2. Leyr is within 25% of the best.
  • Restart, slowest of 5 (ms) (start-up): Leyr 394.7, gonic 393.1. Leyr is within 25% of the best.
  • ping (api latency, one client at a time): Leyr 0.3, Plex 0.2. Leyr is within 25% of the best.
  • search3 (api latency, one client at a time): Leyr 5.4, Navidrome 6.2. Leyr is within 25% of the best.
  • getCoverArt_300 (api latency, one client at a time): Leyr 0.6, Plex 0.3. Leyr is behind the best.
  • ping (api latency, 16 clients at once): Leyr 2.0, Plex 1.9. Leyr is within 25% of the best.
  • getCoverArt_300 (api latency, 16 clients at once): Leyr 2.5, Plex 2.9. Leyr is within 25% of the best.
  • flac_to_mp3_320 (streams: time to first audio byte): Leyr 63.1, gonic 47.1. Leyr is behind the best.
  • flac_to_opus_128 (streams: time to first audio byte): Leyr 62.0, gonic 61.5. Leyr is within 25% of the best.
  • long_flac_original (streams: time to first audio byte): Leyr 1.3, Plex 1.1. Leyr is within 25% of the best.
  • long_flac_mp3_320_from_start_repeated (streams: time to first audio byte): Leyr 59.2, Navidrome 21.2. Leyr is behind the best.
  • long_flac_mp3_320_seek_1500s_repeated (streams: time to first audio byte): Leyr 58.8, Navidrome 8.7. Leyr is behind the best.

Caveats

What these numbers do not show

Noise
The machine is a shared desktop. The 1-minute load average at the start of runs ranged from about 7 to 44, and the same server's concurrent latency changed by up to four times between its two runs. Treat differences under about two times as noise.
Size
One library of about 3,100 tracks per set, one machine, two full runs per server. There are no confidence intervals. Large libraries and many users were not tested.
Different APIs
Jellyfin and Plex have no Subsonic API. Their rows use their closest native requests, which return different data, so they indicate rather than compare. Plex cannot convert audio with no network (see the settings table), so its MP3, Opus and seek cells are n/a. Jellyfin extracted no album art in this setup: its cover-art cells are n/a. Jellyfin did not seek with the request form used: those cells are n/a.
Scan time
It means "the server says it is done and has gone quiet", not "the same work was done". Leyr's time comes from its own log, the others from polling, which adds up to about a second.
Transcoders
Each image ships its own ffmpeg (or, for Plex, its own transcoder). Part of any conversion time is that difference.
Page cache
Music was read from local NVMe and the first run of a server found colder files than later runs. Scan times of the first run are slightly higher for that reason.
The harness
It is Python threads on the same host, so it adds a small, equal cost to every request and cannot go faster than a few thousand requests a second.

Run it yourself

Rebuild the library and test one server

You need Linux, rootless podman, ffmpeg, Python 3.10 or newer, curl and about 35 GB of free disk. The first command sets where the files are downloaded from.

1. Download the scripts

BASE=https://leyr.stream/technology/benchmark/files
mkdir leyr-bench && cd leyr-bench
for f in bench.py report.py make_library.py validate_seek.py; do curl -fsSLO "$BASE/$f"; done
curl -fsSLO "$BASE/public-library-checksums.sha256"

2. Generate the public library

python3 make_library.py ./library --jobs 4 --manifest ./library-checksums.sha256
cmp library-checksums.sha256 public-library-checksums.sha256 && echo "same bytes as ours"

It takes 15 to 25 minutes and about 30 GB. The plan (names, tags, lengths) is identical everywhere. The audio bytes depend on the ffmpeg build; if cmp reports a difference, the generated files still have the same names, tags and lengths. For a quick try, add --albums 10.

3. Test one server

export BENCH_MUSIC="$PWD/library" BENCH_LIBRARY=public
python3 bench.py navidrome --phase main --run 1 --idle 60 --out results
python3 report.py results
less results/report-tables.md

The first command pulls the pinned image, starts a container on an internal network with no route out, scans the library, waits, measures latency and streams, restarts it five times and removes everything it created. Use --idle 300 to match our runs. The other servers are gonic, jellyfin and plex. For Leyr, build an image from the Leyr source with podman build -t bench-leyr . (or set BENCH_LEYR_IMAGE) and run python3 bench.py leyr .... Run it with the machine otherwise idle if you can.

4. Check the seek test

export BENCH_LONG_FLAC_FILE="$(ls library/Test*/*Long*/*.flac)"
python3 validate_seek.py navidrome

A server that starts at the right place prints different loudness for the three offsets, matching the source.

Clean up

podman network rm bench-noegress

Downloads

Scripts, raw results and checksums

Everything is served from this site. Check a download with sha256sum.

Files served from this site.
FileWhat it isSizesha256
bench.pythe harness: one server per run42 kB8c67f21660ab739449d4ec3eb2bd893773117012aff252f2fb8b26800ca5c2dc
rss_breakdown.pysplits a server's resident memory by owner (used by bench.py)4 kB77cb93946660c07038f6fb19cb1a5f4980395b829aa9c73ba9cdeda87052742a
report.pyturns result files into tables9 kB0555f4faeb6a1e687bc9f0c6268b130f480d06d295327fc359345f7bfd61dd11
make_library.pygenerates the public test library12 kB94ea0b32166df6adffd00b2345a9777a31dd401ea87016f06086fe9740c75d09
validate_seek.pychecks that each server really seeks2 kBdd35f3732f3e5414e6cfd91a4fcc8fb845b0efd2688c12b41452ebf0336f1cec
build_subset.shcopies a subset of a music collection (for your own library)1 kB51e6d0152109364ea8a0fe0c6b06ce4207335ee705ffa47bc41d827136757c95
make_page.pybuilds these web pages from the results15 kB55342e98a2551092e78b4726823aa073628c7466a5067cd22f86756bca21909b
run_all.shthe run order used2 kB9be6f6fd45f01e641a6df1e187d84ac3c8bad1bb61398552a84296ea59b2d438
results-public.tar.gzraw result JSON for every run on the public library, with summary.csv, seek-validation.jsonl and report-tables.md24 kBae64e01d05a65694b3976395fc4629fd130c81843bac84b93152213e99d9897a
public-summary.csvpublic library: every value by run11 kB1f12aa184bfd2f8045ed8590afc2d179d043e3b83c2609946aca8c7f6ee0e285
public-seek-validation.jsonlpublic library: the seek check2 kB7d27bce20d219a8ec3a42352d305c520b03e1f2e037a2999907e6aa329171087
results-real.tar.gzraw result JSON for every run on the real library, with summary.csv, seek-validation.jsonl and report-tables.md20 kBfea167f45b990f194fe5787d6b27d2ba4f7725e5e1a803ff588db2d3a6bf1d87
real-summary.csvreal library: every value by run10 kB3e65e1df4a56f5fa1bb7869dcc8bc02cfe39c7cf7fe1bb523d0773f522476e4b
real-seek-validation.jsonlreal library: the seek check2 kB63eb484df3da0b773168512b0ae6bd1c1859b7ef4a392281554a1509eaf9aa53
results-archive-2026-11-13.tar.gzthe first published run (2026-11-13): four servers, the private subset, before the changes listed below22 kB16054f8698890e1107499428d00cc70adddf1280d79e5b7581b9c2d15ed07ab6
archive-2026-11-13-tables.mdthe 2026-11-13 results as a table5 kB101136d6829ed388fb4dddae52efe11c61086fbbe1b86b12ce7bcde84187c583
public-library-checksums.sha256sha256 of every file of the generated library379 kBd4260eb912eda983654021cc37283721d50504b4889a6863abc4bbd82bc30865
public-library-plan.jsonsize and format counts of the generated library0 kB607ae7f2a75625eb1f1b3aeb3c36018205c4f45a99ebab0e2285fda1ced8c39f

Change log

What changed between published runs

Earlier numbers stay available.

2026-11-15
Second published run, on November 14 and 15. Five servers (Plex added), two libraries (the generated public one added; the private one rerun), every container on a network with no route out, Leyr built from the commit above (it includes the memory work and the media tokens), and Leyr's scan time taken from its own log because it finishes before the harness can poll. The public library is the headline. Leyr can be rerun alone; if its rows change, the date above each results table says when each server was measured. Later on November 15, Leyr, gonic and Navidrome were rerun together, one after another, with a Leyr build that keeps transparent huge pages out of the process and returns free memory once the server is quiet; private memory was added next to RSS for every server. Later that evening, Leyr and gonic were rerun together once more, with a Leyr build from before a change to its worker count and purge delay and one after it taken in between, one after another; the before-change runs are in bench/results/memory-333-before in the repository, and Leyr’s private memory under the 16-client load fell from 31–35 MB to 25–30 MB with unchanged request times, as far as a busy host lets us tell.
2026-11-13
First published run: Leyr, Navidrome, gonic and Jellyfin on the private subset only, with network access. Kept as results-archive-2026-11-13.tar.gz and its table. Its stream seek rows for Navidrome and Jellyfin repeated one offset ten times and were partly answered from a cache; the later runs use a different offset per request.