Pure text
On enwik8 zpaq, bsc and PPMd are ahead (19.63%, 20.80% and 21.41% against 21.975%); on Silesia and public documents only zpaq stays ahead, by 1.3 and 1.1 points, and on Silesia it needs twice TSR’s time.
Benchmarks
TSR’s numbers, measured and repeatable
Every ratio is the archive size as a percentage of the original: lower is better. The size of a TSR archive depends only on the files and the profile, so one run is enough, and repeated on another machine it gives the same bytes. Rivals are measured at their best, and placements are stated even when TSR is not first.
The reference corpora
Public sets of files that anyone can download, compressed by TSR at its maximum profile and by ten tools at their maximum levels, on the same machine.
| Corpus | TSR | Best rival | Placement |
|---|---|---|---|
| Office documents (1,322 files, 199.5 MB) | 47.888% | zpaq 51.98% | first of ten |
| Public documents (Govdocs1, 980 files) | 48.598% | zpaq 47.51% | second |
| Silesia (211.9 MB) | 20.124% | zpaq 18.81% | second |
| enwik8 (100 MB of text) | 21.975% | zpaq 19.63% | fourth, behind zpaq, bsc and PPMd |
| PNG graphics (11,356 already-optimised files) | 53.373% | zpaq 52.84% | third |
| Source code (13,663 small files, in a tar) | 14.311% | — | 16.555% on the sum of the files |
| JPEG gallery (5 photographs) | 81.840% | generic ≈ 100% | −18.2%; a specialist about −22% |
The sentence these numbers support: TSR is first on files that already contain compression, and beats 7-Zip, xz, zstd, brotli and bzip2 on text, mixed data and documents. On pure text zpaq, bsc and PPMd stay ahead.
enwik8: ratio and time
A hundred megabytes of text, the corpus on which the strongest compressors are compared. TSR works on independent blocks, one per core: at maximum it takes 19 seconds on all the Mac’s cores, where zpaq takes 106 on a single core and 7-Zip 41.
enwik8, smallest first
- zpaq 19.63% · 106 s
- bsc 20.80% · 4 s
- PPMd 21.41% · 17 s
- TSR maximum 21.975% · 19 s
- 7-Zip 24.80% · 41 s
- xz 24.83% · 53 s
- zstd 25.27% · 51 s
- brotli 25.74% · 123 s
- TSR balanced 28.986% · 3 s
- bzip2 29.01% · 4 s
Nine usage scenarios
Nine folders that resemble a person’s, 3–12 GB each, built from public sources with a licence per file: 44.9 GiB and 195,753 files in all, measured on four machines. Here the Mac with the graphics card off: TSR with the median of three runs, the others at one run at their best.
| Persona | TSR | zip -9 | 7-Zip -mx9 | zstd -19 --long | xz -9 |
|---|---|---|---|---|---|
| Doctor | 34.71 | 67.32 | 46.62 | 45.47 | 57.10 |
| Family | 83.76 | 98.02 | 97.41 | 97.57 | 97.63 |
| Creative | 69.51 | 84.44 | 79.61 | 80.75 | 79.92 |
| Student | 69.39 | 80.35 | 74.71 | 75.07 | 75.19 |
| Legal | 61.91 | 69.98 | 62.07 | 62.76 | 62.50 |
| Videomaker | 97.18 | 99.45 | 98.68 | 98.81 | 98.77 |
| Office | 52.73 | 61.11 | 51.02 | 51.52 | 51.41 |
| IT technician | 54.51 | 57.15 | 51.57 | 51.76 | 54.51 |
| Developer | 43.84 | 52.71 | 36.80 | 38.47 | 39.11 |
TSR is first where the filters work (medical images, phone JPEGs, PNG, PDF), with margins of 5 to 13.7 points over the best of the others. It is behind 7-Zip in solid mode where files are many and small, as for the developer and the office, because it compresses within the blocks of one file and never across files. The balanced profile always makes an archive smaller than zip -9; the fast one always finishes before 7-Zip at level 5, 3 to 38 times sooner.
Where it is behind, and what is missing
Whoever lists only their own advantages forces the reader to find out the rest alone. Here is the rest, with the numbers.
Many small files
On source-code and mail trees 7-Zip in solid mode leads by 2–7 points. It is the price of an archive that opens at any point: TSR compresses within the blocks of one file, never across files.
Music and films
MP3, AAC and already-processed photos: a 1–2% recovery. H.264 films: 2–3.7%. On a video library TSR is worth having for direct access and integrity.
JPEG and PNG
On JPEGs −18.2%, against about −22% for a JPEG-only specialist. On already-optimised PNGs TSR is third.
The cost of maximum
The maximum profile keeps the machine busy for tens of seconds per hundred megabytes per core; on the Ryzen 9 5900X it does 4.0 MiB/s in total. That is why you choose the time.
Code signing
Apple and Microsoft certificates are pending: until then the systems warn on first launch. No certification yet; third-party testing is still to be commissioned.
The method
Four rules
- Sizes are deterministic: one run is enough, and repeated elsewhere it gives the same bytes.
- Times need discipline: where a time supports a claim two alternating runs are compared over three rounds, and the minimum counts.
- Whoever builds the corpus decides the result, so the corpora are built by a script with written criteria: public sources with a verified fingerprint, no file above ten per cent of the mass, no selection by type within a source. An internal corpus used until August 2026 showed a twenty-two-point lead; on corpora rebuilt this way the lead is about three. Corpus withdrawn, measurement redone, episode written down.
- Rivals at their best, including the strongest on ratio alone (zpaq, bsc, PPMd), and negative verdicts published like the positive ones.
The four machines
- Apple M5 Pro, 15 cores, macOS: the time reference
- NVIDIA GB10, twenty ARM cores, Linux arm64
- AMD Ryzen 9 5900X, 12 cores, in Windows 10 and in WSL2 Ubuntu 22.04
The ten tools
7-Zip -mx9 and the PPMd variant (o32 mem2g), zpaq -m5, xz -9e, zstd --ultra -22, brotli -q11, bzip2 -9, bsc -b100 -e2.
On the usage scenarios: zip -9, 7-Zip -mx9 and -mx5, zstd -19 --long=27, xz -9.
Redo it yourself
enwik8 (Large Text Compression Benchmark) and the Silesia Corpus can be downloaded; the rivals run at the maximum levels above. The TSR archive has a deterministic size: one run is enough.
tsr c enwik8.tsr enwik8 --maxtsr gpu-checkTry it on your own files
Beta 0.10.2 is free until 31 January 2027. The archives you create open forever, even after the beta ends.