KDE released Frameworks 6.32 on Friday, 13 November, and the change that matters most to anyone who copies files to USB sticks is that KIO now flushes data to a removable device before it tells the user the copy has finished. All of it is merged, shipped code with a public changelog behind it. None of the speed claims in the release come with published numbers, though.
What KIO changes on removable media
The longest thread in the KDE Frameworks 6.32.0 release announcement and changelog is a run of about ten KIO commits, all tied to bug 281270. KMountPoint can now tell whether a mount is on a removable device. CopyJob uses that to write a copy out before it reports success, to flush the device when a copy fails or is cancelled, and to flush large files that go through the Unix copy path. Moves get a narrower change than a quick summary would suggest: the source is deleted after the flush only when the destination is a removable device, so these commits don’t change a move between two internal disks. The job also shows which source a move is about to delete, and it records a kept source for undo once that source is gone.
In practice, when the progress dialog closes, the bytes should already be on the stick rather than waiting in the page cache for writeback. Data waiting for writeback is why people pull a drive too early and lose a file. The changelog doesn’t measure whether the dialog now stays open noticeably longer on slow flash.
Copy throughput and Samba
Two smaller KIO changes are about speed. copy_file_range calls now line up on 8 MiB boundaries (bug 458001), and a separate commit adjusts the chunk size based on how long each call takes. On the Samba side, the commit that loads Samba usershares with asynchronous processes closes three reports (446923, 513329 and 526668). Another commit stops resolving Samba interfaces when listing usershares, and KDirModel now refreshes folder icons when shares change. On Windows, KIO gains a recycle-bin worker, and the password server no longer waits forever for the wallet.
Baloo, image formats and archives
Baloo’s query layer gets a scope parameter in a Query commit that limits searches to filename, content and similar fields, and baloosearch can now set a default scope. The baloo_file daemon can be started through D-Bus activation, and balooctl uses this to start the daemon after an enable command and restart it after a purge. Content extraction now gets its idle signal from the worker pipe instead of KIdleTime.
KImageFormats adds DDS support for the DX10 block-compressed formats BC4, BC5, BC6H and BC7, decoded with a vendored copy of bcdec, and limits a DDS file to 32 images. The EXR plugin gets the lossy HTJ2K quality setting that OpenEXR 3.5 introduced. The same framework also fixes a JP2 use-after-free, an SGI memory leak and overflow bugs in the IFF and ANI readers, and it now rejects oversized HEIF images before decoding them.
In KArchive, the ZIP code now sets the UTF-8 flag on non-ASCII names. When reading, it handles legacy names in the Windows OEM code page, treats names without the flag as code page 437 and understands the Info-ZIP Unicode Path extra field. Calling these “hardened 7-Zip and TAR readers” overstates the TAR part. K7zip does now check each folder’s coder graph and decodes files on first read instead of on open. The KTar commits, however, add the end-of-archive marker when writing and fall back to the file extension for MIME detection, so they deal with writing and type detection rather than hardening the reader.
Shipped fixes, unpublished timings
Everything here is released, so the gap between what was shown and what was promised is smaller than usual. It shows up in the performance language. The Samba change is presented as a fix for slow listings, and KIO has even added a benchmark for usershare loading, but the announcement gives no before-and-after figures, and the 8 MiB alignment change comes without numbers too. Until someone publishes results from that benchmark, “faster” is a claim, though the mechanism behind it is plausible. The removable-device work has the opposite problem. Its point is correctness, and nobody has publicly measured how much an extra flush costs on slow media.
Requirements and what comes next
Frameworks 6.32 requires Qt 6.10.0. Frameworks comes out monthly, and like the LibreOffice 26.8.3 maintenance release and its 45 bug fixes, it will reach most users through distribution packages. Rolling-release distributions usually get it first, and fixed-release distributions follow on their own schedules. The announcement gives no date for 6.33.
