Download Latest Version zpaqfranz32.exe (9.3 MB)
Home / 65.22
Name Modified Size InfoDownloads / Week
Parent folder
zpaqfranzxp.exe < 2 hours ago 9.1 MB
zpaqfranz-open.exe < 2 hours ago 7.8 MB
zpaqfranzhw.exe < 2 hours ago 8.9 MB
zpaqfranz-full.exe < 2 hours ago 14.7 MB
zpaqfranz32.exe < 2 hours ago 9.3 MB
zpaqfranz.exe < 2 hours ago 9.0 MB
zpaqfranz.cpp < 2 hours ago 11.1 MB
zpaqfranz_solaris < 2 hours ago 7.2 MB
zpaqfranz_oracle79 < 2 hours ago 6.7 MB
zpaqfranz_opensuse < 2 hours ago 7.2 MB
zpaqfranz_openbsd < 2 hours ago 8.7 MB
zpaqfranz_nas_x86_64 < 2 hours ago 7.5 MB
zpaqfranz_nas_i686 < 2 hours ago 8.4 MB
zpaqfranz_macos < 2 hours ago 13.9 MB
zpaqfranz_freebsd < 2 hours ago 7.7 MB
zpaqfranz_fedora < 2 hours ago 7.6 MB
zpaqfranz_esx < 2 hours ago 7.7 MB
zpaqfranz_cortexa72 < 2 hours ago 6.7 MB
zpaqfranz_cortexa57 < 2 hours ago 6.7 MB
zpaqfranz_cortexa55 < 2 hours ago 6.7 MB
zpaqfranz_cortexa53 < 2 hours ago 6.7 MB
zpaqfranz_armv8 < 2 hours ago 6.7 MB
zpaqfranz_armv7_cortexa15 < 2 hours ago 6.8 MB
zpaqfranz_armv7_cortexa9 < 2 hours ago 6.8 MB
zpaqfranz_armv7 < 2 hours ago 6.7 MB
zpaqfranz_armv5 < 2 hours ago 6.9 MB
zpaqfranz_arch < 2 hours ago 8.1 MB
Just about everything source code.tar.gz < 2 hours ago 2.1 MB
Just about everything source code.zip < 2 hours ago 2.1 MB
README.md < 2 hours ago 26.5 kB
Totals: 30 Items   225.7 MB 0

zpaqfranz 65.22 compression levels, and fixes

The theme of 65.22 is ** (a bit) smaller archives, at (almost) no cost**:

  • the default is now -m8l: the same zstd as before, plus its long distance matching. On my PC, 3 to 5% smaller on mixed data, in the same time;
  • a new level, -xl (-m86l): for virtual machines, disk images and big files. It uses more RAM;
  • -compress: a different method for each kind of file (text with -m5, zip and jpg stored, executables with zstd...), optionally by size.

Then many fixes, an issue of -wim images among them (366), and the "kit" of -thunderbird grows: Firefox, Chrome, Edge, Dropbox.

In short:

  • New default -m8l: smaller, same speed, a little more RAM. Every zpaqfranz with zstd, and zpaq 7.15, extract it.
  • New: the letter l in the methods of zstd (-m8l, -m86l, -m7l...): it finds the repetitions far apart in a block, such as the versions of a program.
  • New level -xl = -m86l (64 MB blocks + l).
  • New: -compress: a method for each group of files (text, exe, packed, media...) or extension, and by size. A table at the end says what each group did. The whole story is in compress.md.
  • The last line of a also says how much was written and how big the archive is (all of its parts), even from the window waiting for an elevated run.
  • An image with a -not that left out nothing (issue 366): the archive was left with an incomplete version.
  • -onedrive: the files deleted from the disk were never marked as deleted in the archive.
  • -frugal: the junk of every user of the PC (issue 369); user names with accents; the updater of Chrome is no longer left out of an image.
  • pp and p: CRC-32 too, no more alarms for -wim images, GOOD with folders.
  • New: -firefox, -chrome, -edge, -dropbox, next to -thunderbird (issues 367, 368, 370).
  • Building on Windows: no libraries to link, and the old mingw 7.3 builds again.

Part one, what is new, for the user

  1. The new default: <[inline_block>0](#1
  2. The letter l: long distance matching
  3. A new level: <[inline_block>0](#3
  4. -compress: a method for each kind of file
  5. The last line: what was written, how big the archive is
  6. An image with a -not that finds nothing (issue 366)
  7. -onedrive and the deleted files
  8. -frugal: every user, and what it leaves out (issue 369)
  9. pp and p: CRC-32, -wim images, <[inline_block>3](#9
  10. The kit: -firefox, -chrome, -edge (issues 367, 368)
  11. -dropbox (issue 370)
  12. Building on Windows

1. The new default: -m8l

Up to 65.21 an a with no -m used -m8: zstd at level 3, 16 MB blocks. Very fast, not very small. From 65.22 the default is -m8l: the same zstd, the same blocks, plus the long distance matching of zstd (the l, next section).

My measures, on my PC (16 threads), -m8 against -m8l:

data files size -m8 -m8l time, RAM
c:\zpaqfranz (sources, executables, archives) 9,378 5.5 GB 2,624 MB 2,500 MB (-4.7%) same time; RAM 1.28 => 1.34 GB
c:\msys64 (a compiler, thousands of small files) 116,726 6.2 GB 1,320 MB 1,281 MB (-2.9%) same time
j:\vm\debian12 (the disks of a VM) 16 11 GB 5,440 MB 5,410 MB (-0.55%) +3% time, +3% RAM

With a "mixed bag" of many different files it gets smaller, a little slower and with a little more RAM. With a few big files the difference is small: for those there is -xl (section 3).

And on a Windows 10 VM (8 vCPU), a copy of my working folder: 9,345 files, 5.9 GB, 3.8 GB after deduplication (MiB, seconds, peak RAM in MB from the OS):

method archive a s a RAM t s t RAM
-m8 (the default up to 65.21) 2,498 11.3 785 4.6 322
-m8l (the default now) 2,380 11.2 785 4.8 293
-m86l (-xl) 2,174 11.2 2,396 4.7 1,044
-m9 (-good) 2,318 13.7 4.5
-m7 (-better) 1,998 157 3,072 4.3 1,055
-m5 (-max) 1,947 1,808 1,826
  • Extraction does not change. An -m8l block is a normal zstd frame: every zpaqfranz with zstd decodes it (I tried 65.3 and 65.22b, both older than the l), and so does zpaq 7.15, through the ZPAQL decoder stored in the archive.
  • -normal is -m8l too (it was the same as the default, and stays so). The default of repack is -m8l.
  • An archive made with no -m is not the same, to the byte, as one of 65.21. With -m8 written it is.
  • On ESXi and on the NAS builds (the a of zpaq 7.15) the default stays -m1, as before. Without zstd (-715, the builds without it) too.

2. The letter l: long distance matching

l, written before the first comma of a zstd method, turns on the long distance matching of zstd:

-m8l       level 3, 16 MB blocks (the default)
-m86l      level 3, 64 MB blocks (-xl)
-m7l       = -m86lh19
-m9l       = -m84lh9
-m86lh9    64 MB blocks, level 9
-m8a3l     fast level -3

It works in -compress too (exe:86l).

Why. zpaqfranz gives zstd a window as big as the block (16 or 64 MB), but at low levels the match finder of zstd keeps small tables and looks near: it does not see a repetition many MB back. The versions of the same program, the copies of a file inside a block, the same structures in a disk image are just that. The long distance matching is a second match finder, made for long repetitions far apart. It costs little time, because it looks only for long ones.

  • It is not always smaller. With 16 MB blocks it can gain little, or even lose a little: -m9l was 0.15% bigger on a small folder (its matches took the place of better ones). This is why -good stays -m9, and -fast stays -m80.
  • It shines with big blocks: on the folder above -m86l is 13% smaller than -m8, in the same time.
  • -turbo, -noturbo and -t1 make the same archive, to the byte, as always.

3. A new level: -xl

-xl = -m86l: 64 MB blocks and the long distance matching. For virtual machines, disk images and big files: their repetitions are far apart, and a 64 MB window finds them.

zpaqfranz a z:\vm.zpaq j:\vm\debian12 -xl

It uses more RAM: on the folder above about 3 times -m8l, both in a (2.4 GB against 0.8) and in t (1.0 GB against 0.3). Less than -m7 in a, the same in t. On a small NAS, or a 32 bit system, stay with the default.

The levels by name, now:

level method what
-lean -m80 -t2 -noturbo minimal resources, slower (a NAS)
-fast -m80 a bit faster and bigger than -normal
-normal -m8l balanced speed and size: the default
-good -m9 a bit smaller and slower than -normal
-xl -m86l VMs, disk images, big files (more RAM)
-better -m7 high compression, fast extraction
-ultra -m6 very slow compression and extraction
-max -m5 extreme compression, extremely slow

One level only, and not with -m (03995, exit code 2). Without zstd, -xl is -m2.

4. -compress: a method for each kind of file

A single -m is a compromise. It spends CPU on JPGs and ZIPs that cannot shrink any more, and compresses source code, which would get much smaller with a strong model, at the same level as everything else. With -compress every group of files gets its own method:

zpaqfranz a z:\1.zpaq c:\data -compress "packed:0;media:0"
    as usual, but no time lost on zip, jpg, mp4... (stored, still deduplicated)

zpaqfranz a z:\1.zpaq c:\data -compress "text:5;exe:7;packed:0;media:0"
    text with the strongest model, executables with zstd, compressed files stored

zpaqfranz a z:\1.zpaq c:\data -compress "text:5/20MB:7"
    text up to 20 MB with -m5, bigger text files (dumps, logs) with -m7

zpaqfranz a z:\1.zpaq c:\data -m5 -compress "*:/1GB:8"
    everything at -m5, except the files over 1 GB: those with -m8
  • Keys: one of eight groups (text exe raw db office disk packed media, by what is inside the files: a .docx is a zip, so packed), an extension (it wins over its group), none (no extension), * (every other file). What no key takes gets the -m of the command.
  • Methods: any method of -m, written whole (5, 7, 86h19, 86l, x4.3ci1). A wrong one is an error at once, before anything is read.
  • By size: key:method/size:method/..., the size included; K M G T x1000, KB MB GB TB x1024. A step with no method is "as if the key were not there".
  • -compressgroup name:ext,ext makes a group again, or a new one. In a -ini too.
  • A block never holds two groups, so every block has its own method. The archive is a normal zpaq archive: x, t, mount and zpaq 7.15 read it as always.
  • At the end, a table: files, bytes read, new bytes after deduplication, compressed, ratio, blocks and CPU seconds of each group. Change a group, run again, compare.

How much does it matter? On the folder of the VM above, the best mix (-m9 -compress "text:7;exe:7;none:7;packed:8;media:0") made 2,055 MiB in 118 s: about 30% faster than -m7 (1,998 MiB, 171 s) for an archive 3% bigger, and against -m5 (1,947 MiB) 15 times faster to add and 400 times faster to test. The archive was not smaller than with a single well chosen -m: the gain is time, no CPU spent where it gives nothing.

Not on ESXi and the NAS (they use the a of zpaq 7.15), not with -image, -stdin, -m s. Everything else, with examples, the numbers, recipes and how to test: compress.md.

5. The last line: what was written, how big the archive is

The last line of an a now also says what this run added to the archive, and how big the archive is on the disk:

0.125s (00:00:00,50.03MB,OS REPORT 197.96MB, WRITTEN 6.73MB, ARCHIVE 6.73MB) (all OK)
0.015s (00:00:00,21.96MB,OS REPORT 18.59MB, WRITTEN 35.72KB, ARCHIVE 6.77MB) (all OK)
  • ARCHIVE is the whole archive: with a multipart one ("z:\pc_????.zpaq"), every part, the part 0 of -index included; with -chunk, every chunk; with Franzen, the .franzen too.
  • WRITTEN is how much it grew in this run (an update with nothing new: a few bytes, the empty version).
  • Also from the window that waits for an elevated run (a -vss launched by a standard user, "getting the power!"): that window does not see what the other one writes, but it measures the archive before and after.
  • Not there with a "" (no archive), nor for the commands that write no archive.

6. An image with a -not that finds nothing (issue 366)

Windows. An image (ac, or a X: -image) with a -not that left out no file ended with

array_to_ramfile of empty vector!

and the archive was left with an incomplete version (with -vss, a shadow copy was left behind too). The report came with -not :X (a typo for X:) and -wim, from WinPE, but neither mattered: any -not that matched nothing, even -not D:/nothere/, did it.

Now the image is made, and every -not of an image that leaves out nothing is said, in yellow, always (before: only with -verbose). This is what would have shown the typo:

-not <<:X>>: no such file on C: (nothing left out for it)

The exit code does not change for these lines.

Thanks to Erol-2022.

7. -onedrive and the deleted files

Windows. With -onedrive (the files that are only in the cloud are not downloaded) a file deleted from the disk was never marked as deleted in the archive: the next version still had it, forever ("0 -removed"). To tell a placeholder of OneDrive from a real file zpaqfranz looked at the file on the disk, and a file that is not there any more was taken for a placeholder, which must never be touched.

Now a file that is not on the disk is deleted, as without -onedrive; a placeholder stays as it was (its last version in the archive is not touched). At the end of a, a line says how many online-only files are not in the archive (67113).

8. -frugal: every user, and what it leaves out (issue 369)

Windows. -frugal leaves out the junk: caches, temporary files, thumbnails, crash dumps.

Every user (issue 369). The junk of the user folders was taken only for the user running zpaqfranz. From a scheduled task (SYSTEM), or as an administrator other than the user of the PC, the 2.7 GB of the thumbnail cache of the real user went into every image. Now:

  • a X: -image -frugal and q -frugal: the junk of every user of the PC (from the registry, as -thunderbird -all). The users whose folders cannot be read are said (94111): it takes an elevated administrator;
  • work frugal -all: the same, to see it (work frugal alone is still the user running).

What else changed

  • User names outside the ANSI code page (José, Иван): their junk was never left out. Now the names are UTF-8, as in the rest of zpaqfranz.
  • An image of D: with -frugal printed a yellow line for every junk folder on C: (68 on my VM). Now those are left alone; one line says how many, with -verbose.
  • The updater of Chrome (Google\GoogleUpdater, and the same program for Dropbox, DropboxUpdater) was left out whole. An image restored had no updater: Chrome stopped updating itself, and only reinstalling it fixed that. Now only its download cache (crx_cache, 551 of 562 MB on my VM) and its crash dumps are left out.
  • Dropbox: not the whole sync folder, only its block cache (local_block_cache). The rest is the state of the synchronization.
  • -kill (which deletes the junk) skips the index of Windows Search, held open by the service. In images and backups it is still left out: Windows rebuilds it.
  • -thunderbird -frugal: also the telemetry (datareporting, saved-telemetry-pings, Pending Pings), issue 367.

Thanks to 8tpercent, for issues 367, 368, 369 and 370.

9. pp and p: CRC-32, -wim images, GOOD

pp (the paranoid test for archives on SSD: every file rebuilt, its hash computed again) now checks the CRC-32 of every file too, not only its hash. Two columns, hash and CRC-32, a row only when it is not zero, a color for each one. The CRC-32 is almost free (0.08 s on 400 MB): pp is slow because of the decompression. With -verbose, the time it took.

  • -wim images. An image made with -wim and blocks skipped has no hash: the blocks not read cannot be hashed. Only its CRC-32 can be checked. pp said WARN: some files does not have stored hash, in a way that scared. Now it is a note (05564), the CRC-32 is checked, and the result is GOOD when it is right.
  • GOOD never came out with a folder in the archive (the folders were counted as files). Now it does.
  • p checks by CRC-32 the files with no hash (05569): a -wim image went from exit code 2 to 0.
  • p and a wrong hash: it was an error only with -verify (without it, "All OK"). I found it reading the code; I could not make a real case (a damaged archive stops p earlier).
  • pp: a wrong CRC-32 is exit code 2 (05565).

10. The kit: -firefox, -chrome, -edge (issues 367, 368)

Windows (not in the open build). -thunderbird of 65.20 moved the mail from a PC to another. Now the browsers too, with the same rules, alone or together:

zpaqfranz a z:\kit.zpaq -thunderbird -firefox -chrome -edge -all -kill -frugal
           every user of the PC (an elevated administrator), the programs closed first, no junk

zpaqfranz x z:\kit.zpaq -thunderbird -firefox -chrome -edge
zpaqfranz x z:\kit.zpaq -firefox -to caio        into another user (one user in the archive)
  • Every profile, by name. Firefox and Thunderbird from profiles.ini, Chrome and Edge from their Local State: each one is said with the name you see in the browser. Home, office, husband, wife: all of them.
  • -frugal, for each program: its caches, crash reports, telemetry, temporary files, and for Chrome and Edge about thirty components they download again by themselves (Edge on my VM: from 135.6 to 10.9 MB).
  • x never destroys what is there: as with Thunderbird, it is renamed (Firefox.001, User Data.001) and then the profiles are extracted.
  • Firefox (or Thunderbird) installed in another folder did not know the restored profiles, and made a new empty one. Now the restored profiles are given to that installation too (67105).
  • Chrome and Edge, passwords and cookies. Windows keeps their key, for that user on that PC (DPAPI, and from Chrome 127 the "app-bound encryption"). On the same PC and the same user they come back. On another PC, or into another user, they are lost: bookmarks, history, extensions and settings come back anyway. zpaqfranz says it, at the backup and at the restore (67100, 67101, 67106), and says how to export the passwords first. It does not touch those keys: that is what the programs that steal passwords do.

Tried for real: a kit made on a Windows 11 VM restored into another PC (a Windows 10 VM): the profiles, Firefox and its installation open; Chrome and Edge say in their log that they cannot decrypt the key, as expected.

11. -dropbox (issue 370)

Windows. The folders of Dropbox, found by themselves (its info.json, every account), as -onedrive does for OneDrive:

zpaqfranz a z:\dropbox.zpaq -dropbox                    the user running, online-only files not downloaded
zpaqfranz a z:\dropbox.zpaq -dropbox -all -download     every user, the online-only files too
zpaqfranz x z:\dropbox.zpaq -dropbox -to z:\restored     into a folder (not synced)
  • Online-only files are not downloaded, as with -onedrive; a line says how many were left out. With -download they are downloaded while they are read (then they stay on the PC), after checking the free space (67116, exit code 2 if it is not enough). Dropbox must be running (67117).
  • .dropbox.cache (old and deleted files Dropbox keeps) is left out, always.
  • x -dropbox puts them back where they were, or -to a folder or a user. Nothing is renamed: a Dropbox folder renamed is a folder deleted in the cloud. So x -dropbox goes alone, not with the other switches of the kit (67120).
  • Into a folder synced with the cloud (Dropbox, OneDrive, any provider): a captcha. What is written there goes up to the cloud, to every PC and everyone sharing it (67130).

-virtualbox, from the same issue: not now.

12. Building on Windows

  • No libraries to link. g++ -O3 zpaqfranz.cpp -o zpaqfranz.exe builds it (add -static for an exe that needs no DLL of the compiler). The download (urlmon.dll) is loaded at run time, as wininet.dll already was: no more -lurlmon. Not -O0: on Windows the object file gets too many sections (or add -Wa,-mbig-obj).
  • The old mingw 7.3 builds again (from 65.5h it did not: vsnprintf_s in Mbed TLS).
  • gcc 16: the 7 warnings with the options of my build are gone.
  • The source is now about 290,000 lines, too many for the column tracking of gcc: a note (not a warning) says so. -flarge-source-files makes it go away.
  • The examples of build lines in the source, for Windows and for every *nix, are the ones I really use now (the old ones are kept, below them).

Part two, the details, for power users

1. What can change for a script

  • a with no -m: -m8l instead of -m8. A different archive (smaller), a little more RAM. The same for -normal and for repack with no -m. On ESXi, the NAS and without zstd: -m1, as before.
  • The last line of a: , WRITTEN x, ARCHIVE y inside the parentheses, after OS REPORT (with an archive only). (all OK) is still at the end.
  • An image with a -not that leaves out nothing: a yellow line, always (before: only with -verbose). The exit code does not change. And the image is made (before: an incomplete version, exit code 2).
  • -onedrive: the deleted files are marked as deleted (before: never).
  • -image -frugal, q -frugal: the junk of every user (before: of the user running). The images can be smaller.
  • -frugal: the updaters of Chrome and Dropbox are kept, only their caches are left out: an image can be a few MB bigger.
  • pp: GOOD with folders too; a wrong CRC-32 is exit code 2; a -wim image with no hash is a note, not a warning.
  • p: a -wim image with no hash is exit code 0 (before: 2); a wrong hash is an error without -verify too.
  • -compress with -image, -stdin or -m s: exit code 2 (67149). On ESXi and NAS: exit code 2 (67147).
  • -xl with -m, or with another level: exit code 2 (03995).

2. The switches

Switch What
-m8l, -m86l, -m7l, -m9l, -m86lh9... zstd with long distance matching (the l before the first comma)
-xl the level -m86l: VMs, disk images, big files (more RAM)
-normal now -m8l (as the default)
-compress X a method for each group of files: key:method;key:method, keys = group, extension, none, *; by size key:method/size:method
-compressgroup X a group made again, or a new one: name:ext,ext,... (they add up, also in a -ini)
a -firefox, a -chrome, a -edge (Windows) the browser of the user, every profile; -all -kill -frugal as -thunderbird
x -firefox, x -chrome, x -edge (Windows) back: what is there renamed .001; -to user
a -dropbox (Windows) the Dropbox folders of the user, online-only files not downloaded; -all every user
a -dropbox -download the online-only files too, downloaded while read
x -dropbox back where they were, -to a folder or a user; a captcha into a synced folder
work frugal -all (Windows) the junk of every user of the PC

3. The messages

Code When
03995! a level by name: one only, and not with -m (now with -xl in the list)
05564 pp, p: images of -wim with no hash, normal (their CRC-32 checked)
05565 pp: a wrong CRC-32
05566-05568 pp -verbose: files with no CRC-32, CRC-32 different, the time of the CRC-32
05569 p: how many files with no hash were checked by their CRC-32
44121$ 44122$ 44124$ 44125$ 44151$ 44154$ a -not of an image that leaves out nothing (always, not only with -verbose)
66887 the kit: the profiles of a program of a user, by name
67100, 67101, 67106$ Chrome, Edge: passwords and cookies only on the same PC and user
67102$, 67103, 67104, 67107$ the kit in x: what the archive has, and what it has not
67105 an installation of Firefox or Thunderbird in another folder, given the restored profiles
67110-67118 a -dropbox
67120!, 67129!-67131! x -dropbox
67140!-67156! -compress, -compressgroup: the errors, what goes where, the table at the end
94110, 94111$, 94112, 94114 -frugal: the users taken, the ones that cannot be read, the folders on other drives, Windows Search

Part three, for developers

1. Tried, and left out: a level by compressibility

zpaqfranz estimates how much each block can be compressed, before compressing it. The idea was to use that number to raise the level of the blocks that compress well, by themselves (never up to -m5: too slow). I measured it, block by block, on the folder above: the estimate tells where the gain is, but the cost is in the same place (zstd spends little on data that do not compress, at every level). -m9 only on the blocks that compress made -6.97% for 52 s of CPU; -m9 everywhere -7.24% for 56 s. Nothing to gain: left out. The l came from the same measures.

2. Known limits

  • The l is not always smaller: with 16 MB blocks it can lose a few tenths of a percent (-m9l). On my data, at level 3 (-m8l, the default), it was always smaller.
  • -xl: about 3 times the RAM of -m8l. On the NAS builds (32 bit i686 too) it was tried on a small folder only: a, t, x right, the RAM on big data not measured.
  • -compress is not in the a of zpaq 7.15 (ESXi, NAS). The archive is not smaller than with a single well chosen -m: the gain is time.
  • Chrome and Edge: passwords and cookies come back only on the same PC, into the same user. Not tried: Chrome of many real users with -all.
  • -dropbox: on my VM the "online-only" files were in fact on the disk. Not tried: really online-only files, -download that downloads them, -vss with placeholders, x into a synced folder accepted (it would write into a real cloud).
  • Issue 366: the same fix is in the restore of an image (xc), read, not run (it writes a whole disk).

Download zpaqfranz

Source: README.md, updated 2026-11-19