Where does the compression ratio end

The common archiver is already good enough for almost everything, and yet somebody always wants to know where the limit actually sits. PAQ8px belongs to the PAQ family, a line of experimental compressors built around context mixing, and it does not chase practicality. It chases one thing: making the output a byte smaller. In public compression benchmarks the PAQ algorithms sit at the top of the ratio tables year after year, which makes them the reference point that extreme compression research has to reckon with.

Trading time for size

The principle is not complicated to describe. Several dozen prediction models, each tuned to a different kind of data, run at the same time, and a mixer in the style of a small neural network combines their predictions with weights that rise as a model proves accurate. Text, images and audio have their own models, and even the text stream inside a PDF can be recognised and handled separately.

The price is stated just as plainly. At a high compression level a single file may take minutes, and memory use ranges from a few hundred MB to nearly 30 GB depending on the setting. The project itself says this is a research tool and not a candidate for a production environment. The command line interface and the versioned suffix on the output name carry the same academic manner.

Why it exists at all

The source is published under the GPL on GitHub, and the pace of work has continued through 2026, with the v216 release arriving at the end of 6 month and the image and audio models still being refined. Three things make it useful. It can estimate the entropy ceiling of a file, which tells you how much room is left. It serves as a baseline for compression research. And it satisfies the challenge of taking 1GB down to 200MB, which is a question people ask even when the answer has no practical use.

For daily backups the ordinary archiver remains the right answer, and nothing here changes that. The reason to run this one at least once is different: to see where the boundary of the possible is currently drawn.

Reading the result honestly

A ratio table on its own can mislead. The gain over a good general purpose compressor on ordinary documents is small in absolute terms, and it is paid for in minutes of processor time and gigabytes of memory. Where the difference becomes interesting is in material with structure that general compressors do not model, which is exactly why specialised models exist inside this program.

That is also the honest summary of the whole family: a laboratory instrument rather than a tool in the drawer, useful mostly to people who want to know what the ceiling looks like.

What to do with it in practice

A sensible first experiment is a small file whose contents you already know, run at a moderate level so that the result arrives while you are still interested. Compare the output size against the archiver you normally use, and look at how long the run took. That comparison, rather than any table published elsewhere, is the one that says whether the trade is worth making for your data.

The second experiment is a file you cannot compress well at all, such as one that is already packed. If the output comes out slightly larger than the input, the run is telling you something useful about the entropy of your material, and that is a result rather than a failure.

This is the kind of program that hardly anybody uses daily while the whole field stands on its shoulders. It pushes the mathematical limit of compression outwards again and again, and the mainstream archivers then pick the practical parts out of its papers and its code. Being slow is its temperament, being strong is its position.