Three publishable runs under the exclusive bench queue, one per filtered group, 3 repetitions (tools 3), run on the build of 2026-11-02 (main 33c568d plus the bench rows). Groups: A mp4-mux, mp4-remux, mkv-mux; B tagw-id3, tagw-flac, tagw-ilst, tagw-riff, tagw-mkv; C remux-mem, playable. Comparators: SharpMP4 0.3.2, TagLib# 2.3.0, ATL.NET 7.19. Method: docs/bench.md.

Against the bar

Bar: geometric mean of the ratios <= 1.00x and no file worse than 1.25x, against the fastest managed comparator per file.

Notes

• Each group ran as its own job under the exclusive benchmark queue, alone on the machine. Load before/during/after is in each group's header below; the load average also counts the shell steps around the run.
• Tag-write rows edit the same four common fields (title, artist, album, comment) of a file held in memory; method and the difference in work per implementation are in docs/bench.md. The ratio is against TagLib#, the fastest managed comparator on every file here; ATL.NET is 1.2x to 18x slower than TagLib# (its ratio column), least on WAV and most on APE-tagged MP3.
• tagw-mkv is not a comparison of breadth: neither comparator is a Matroska writer in practice. ATL.NET fails on both files (NullReferenceException) and TagLib# fails on test1.mkv ("Invalid EBML format"), so the row has one file, h264-two-audio-tags.mkv, where TagLib# reads and writes it. The ratio there is 0.99x (OneMedia 121.8 us, TagLib# 123.4 us) and the bar is met on this machine.
• Rows mkv-mux, remux-mem and playable have no managed comparator (no managed Matroska writer; no managed remuxer), so they pass; the times next to ffmpeg 7.0.2 and mkvmerge 103.0 are information. mkvmerge exits 1 on these files (warnings about the input) and its time is still the whole run.
• remux-mem: OneMedia remuxing the two files over 4 GiB (a 4.08 GiB MP4 with the movie box at the end, a 4.14 GiB Matroska file) into MP4 and Matroska on disk peaks at 62 to 95 MiB resident, process start-up included (baseline about 29 MiB); ffmpeg -c copy to Matroska peaks at 31 MiB. Allocation per run is 2 to 25 MiB.
• playable (MakePlayable for the generic MP4 profile on a 0.2 MiB Matroska file) takes 160 ms here against 7.6 ms for ffmpeg (information). An indicative run on the same machine of the build before the writer fixes took 9.6 ms; the difference is in what the new save path does (flushing the temporary file to disk before the rename), which a 0.2 MiB file notices and a large one does not. It is reported, not tuned.

Group A: mp4-mux, mp4-remux, mkv-mux

Run 2026-11-02 04:44Z

• Machine: AMD Ryzen 9 7950X3D 16-Core Processor; 32 logical cores; Linux Mint 22.3 X64
• ISA: sse4.2 avx2 bmi2 avx512f avx512bw
• Storage of the corpus: /dev/nvme0n1p4; every file is read once before its rows, so it is in the page cache for all of them
• .NET: 10.0.13 (.NET 10.0.13)
• Load average (1 min): before 2.66; during mean 2.26, max 2.66 over 10 samples at 5 s; after 1.88
• Elapsed: 0.8 min; repetitions 3 (managed), 3 (tools); files over 1 GiB get one warm-up run and 3 samples of one run each, and tools once
• Packages as restored: SharpMP4 0.3.2; RecursiveAI.OneAudio 1.6.1; MetadataExtractor 2.9.3; TagLibSharp 2.3.0.0; z440.atl.core 7.19; Matroska 0.0.12
• Tools: ffmpeg version 7.0.2-static https://johnvansickle.com/ffmpeg/ Copyrig; MP4Box - GPAC version 26.07-rev0-ga07cbfff-master; mkvmerge v103.0 ('The Darkness') 64-bit; MediaInfo MediaInfo Command line,; ExifTool 13.60

Ratio is OneMedia's median over the fastest managed comparator's median on the same file (below 1.00× is faster). Peak is the process's peak resident set; Baseline is its resident set before any work, so Peak − Baseline is what the work added.

mp4-mux: MP4: write the first AAC track's frames from memory to an MP4 in memory

mp4-remux: MP4: read the first AAC track from a file and write it as an MP4 in memory

Reference tools, run as processes (wall time including start-up, peak resident set); information only:

mkv-mux: Matroska/WebM: read every packet and write it to a new Matroska file (no managed Matroska writer to compare with)

Reference tools, run as processes (wall time including start-up, peak resident set); information only:

Group B: tagw-id3, tagw-flac, tagw-ilst, tagw-riff, tagw-mkv

Run 2026-11-02 04:45Z

• Machine: AMD Ryzen 9 7950X3D 16-Core Processor; 32 logical cores; Linux Mint 22.3 X64
• ISA: sse4.2 avx2 bmi2 avx512f avx512bw
• Storage of the corpus: /dev/nvme0n1p4; every file is read once before its rows, so it is in the page cache for all of them
• .NET: 10.0.13 (.NET 10.0.13)
• Load average (1 min): before 1.73; during mean 1.45, max 1.73 over 14 samples at 5 s; after 1.30
• Elapsed: 1.1 min; repetitions 3 (managed), 3 (tools); files over 1 GiB get one warm-up run and 3 samples of one run each, and tools once
• Packages as restored: SharpMP4 0.3.2; RecursiveAI.OneAudio 1.6.1; MetadataExtractor 2.9.3; TagLibSharp 2.3.0.0; z440.atl.core 7.19; Matroska 0.0.12
• Tools: ffmpeg version 7.0.2-static https://johnvansickle.com/ffmpeg/ Copyrig; MP4Box - GPAC version 26.07-rev0-ga07cbfff-master; mkvmerge v103.0 ('The Darkness') 64-bit; MediaInfo MediaInfo Command line,; ExifTool 13.60

Ratio is OneMedia's median over the fastest managed comparator's median on the same file (below 1.00× is faster). Peak is the process's peak resident set; Baseline is its resident set before any work, so Peak − Baseline is what the work added.

tagw-id3: MP3: change title, artist, album and comment; the edited file in memory

tagw-flac: FLAC: change title, artist, album and comment; the edited file in memory

tagw-ilst: MP4/M4A: change title, artist, album and comment in ilst; the edited file in memory

tagw-riff: WAV: change LIST INFO title, artist, album and comment; the edited file in memory

tagw-mkv: Matroska: change title, artist, album and comment; the edited file in memory

Group C: remux-mem, playable

Run 2026-11-02 04:46Z

• Machine: AMD Ryzen 9 7950X3D 16-Core Processor; 32 logical cores; Linux Mint 22.3 X64
• ISA: sse4.2 avx2 bmi2 avx512f avx512bw
• Storage of the corpus: /dev/nvme0n1p4; every file is read once before its rows, so it is in the page cache for all of them
• .NET: 10.0.13 (.NET 10.0.13)
• Load average (1 min): before 1.20; during mean 1.17, max 1.26 over 15 samples at 5 s; after 1.17
• Elapsed: 1.2 min; repetitions 3 (managed), 3 (tools); files over 1 GiB get one warm-up run and 3 samples of one run each, and tools once
• Packages as restored: SharpMP4 0.3.2; RecursiveAI.OneAudio 1.6.1; MetadataExtractor 2.9.3; TagLibSharp 2.3.0.0; z440.atl.core 7.19; Matroska 0.0.12
• Tools: ffmpeg version 7.0.2-static https://johnvansickle.com/ffmpeg/ Copyrig; MP4Box - GPAC version 26.07-rev0-ga07cbfff-master; mkvmerge v103.0 ('The Darkness') 64-bit; MediaInfo MediaInfo Command line,; ExifTool 13.60

Ratio is OneMedia's median over the fastest managed comparator's median on the same file (below 1.00× is faster). Peak is the process's peak resident set; Baseline is its resident set before any work, so Peak − Baseline is what the work added.

remux-mem: Remux of the files past 4 GiB into MP4 and Matroska, output to disk: peak resident memory

Reference tools, run as processes (wall time including start-up, peak resident set); information only:

playable: MakePlayable (generic MP4): Matroska H.264+AAC to MP4; information

Reference tools, run as processes (wall time including start-up, peak resident set); information only: