Linux Kernel 7.3: The Release That Wants to Be Fast, and Is Fighting About AI
Linux 7.3 is in release candidate and heading for an October 2026 launch, and it may become the next long term support kernel. It is one of the most performance focused releases in years, with big Btrfs and EXT4 speedups, memory and scheduler fixes, initial Apple M3 boot support, and a community openly split over AI written patches. Here is what to expect and where the main distributions stand.
📅
✍️ Gianluca
Linux Kernel 7.3: The Release That Wants to Be Fast, and Is Fighting About AI
While most people still picture Linux as a 6 point something kernel, the project quietly moved on. Linux 7.0 shipped in April 2026 and became the default kernel in Ubuntu 26.04 LTS. Then came 7.1 and 7.2, and as of mid September 2026 we are watching 7.3 take shape in the release candidates, with the stable version expected in the second half of October. It is early enough that details can still shift, but the shape of this release is already clear, and it is an unusually interesting one.
Two things make 7.3 worth paying attention to now rather than later. First, it is being described as one of the most performance focused kernels in years, and the benchmarks behind that claim are real, not marketing. Second, it lands in the middle of the loudest argument the kernel community has had in a long time, and the argument is about artificial intelligence writing kernel code. And there is a practical reason to care beyond curiosity: 7.3 is a likely long term support candidate, which means it is probably the kernel your distribution will carry for years.
The 7.x Era Arrived Without Much Ceremony
The jump to version 7 was not a rewrite or a break in compatibility. Linus Torvalds bumps the major number when the minor number climbs high enough that he stops enjoying the arithmetic, exactly as he did going from 4 to 5 and from 5 to 6. So 7.0 was, in engineering terms, just the release after 6 point something big. What is new is that the 7.x line is now the mainstream, already the default in the newest Ubuntu long term support release, tracked by RHEL 10 and its rebuilds, and rolling out everywhere else. 7.3 is the third step in that line, and the most feature dense so far.
A Release Obsessed With Speed
The headline for 7.3 is performance, and unusually the gains are spread across the parts of the kernel that touch almost everyone: storage, memory and scheduling. These are not synthetic peak numbers, they come from the pull requests and independent benchmarking, and several of them are the kind of fix where someone found a decade old inefficiency and simply removed it.
Linux 7.3, by the Numbers
Btrfs gets some of the biggest wins. Two separate fixes in the logging and fsync paths add roughly 5x each, one by removing an unnecessary delay when several logging tasks run at once, the other by skipping pointless hole detection during a full fsync on files that have many extents but no holes. Direct I/O throughput on the same filesystem climbs from around half of the theoretical maximum to roughly 95 percent.
EXT4 is not left out. A rework of the IOmap layer pushes it from about 1.92 million to 2.19 million IOPS on fast NVMe in testing, and parallel direct I/O writes are finally unlocked where it is safe to allow them.
The quietest but most striking fix is in memory management. Engineers at ZTE tracked down a bug in the kernel same page merging code where an application could freeze for hundreds of milliseconds under memory pressure. The fix drops the worst case lock hold time from about 705 ms to a 1.44 ms average. On top of that, compressed memory through zsmalloc runs meaningfully faster, up to 1.83x on a Raspberry Pi 4B.
Scheduling gets attention too. A set of changes reworks how the scheduler picks tasks across control groups, smoothing frame rates and consistency on gaming hardware, and there are cluster aware fixes for Intel hybrid chips with a mix of performance and efficiency cores. For latency sensitive workloads, a networking test with DPDK saw the worst case delay fall from about 17 ms to roughly 1.5 ms.
There is also a milestone for the pluggable scheduler framework. With 7.3, the sub scheduler support in sched_ext is now considered feature complete. In plain terms, a scheduler written in BPF can hand an entire control group subtree to a nested scheduler, granting and later revoking per CPU capabilities like the right to enqueue tasks or control frequency. That is a serious foundation for people who want to experiment with custom scheduling policies without patching and rebuilding the kernel, and it arrives with the safety work to match, including making the abort path safe under non maskable interrupts.
The Hardware: Apple M3, AMD Zen 6, and Yes, a Steam Controller
On the hardware side, the most eye catching addition is initial mainline support for Apple M3 Pro, Max and Ultra chips. It is important to be honest about what that means today: this is initial boot support through device tree, enough to start the machine on a stock kernel, not a finished desktop. There is no GPU acceleration yet and plenty is still missing. It is a foundation, and a meaningful one for anyone who wants Linux on recent Apple Silicon without relying entirely on out of tree work, but it is not the day you replace macOS.
Elsewhere the list is long and practical: enablement for AMD Zen 6, new AMD UALink interconnect and CXL Type 2 accelerator support aimed at the data center, a driver for the RGB lighting on Ryzen AI Halo systems, eleven new instruction set extensions for RISC-V, Thunderbolt on Apple Silicon, better Intel thermal handling, and a compatibility fix that lets the in kernel SMB server work with Apple Time Machine backups. There is even initial driver support for the 2026 Steam Controller, which is a small thing that tells you how broad the contributor base still is.
The Fight Nobody Could Avoid: AI Inside the Kernel
The technical story is only half of what makes this cycle notable. The other half is a genuine community fight, and it is about whether and how artificial intelligence belongs in kernel development. It came to a head in 2026 on the mailing list, where Torvalds pushed back hard against developers who wanted the project to take an openly hostile stance toward AI.
Fork It or Move On
In July 2026 Torvalds told the mailing list that "Linux is not one of those anti-AI projects", framing AI as just a tool and, in his words, "clearly a useful one", while conceding it "can also be a somewhat painful tool". His message to people who wanted a blanket ban was blunt: fork the project or move on. The existing policy already allows AI assisted contributions, but with a firm condition. A human has to review the code, disclose that a tool helped produce it, and formally sign off on it. The machine does not get a Signed off by line of its own.
The friction is not theoretical. Google engineers introduced an agentic code review system into the kernel workflow that uses large language models to hunt for bugs in patches, and separately a wave of AI driven bug bots began flooding maintainers with reports, many of them against ancient network drivers that almost nobody still runs. Maintainers ended up spending hours deciding whether machine generated findings were real, already known, or simply noise. That is the honest tension of 7.3: some of the same automation that finds real security flaws also generates a firehose of work.
Whatever your opinion of AI, this is a healthier debate than the headlines suggest. The kernel did not ban the tools and did not surrender to them either. It kept the one rule that has governed the project for decades: a named human is accountable for every line that goes in. That is a more durable answer than most industries have managed.
Where the Distributions Stand
If you are wondering when 7.3 reaches you, it depends entirely on the kind of distribution you run, and right now most of the ecosystem is still catching up to earlier 7.x releases.
Rolling release distributions are the fastest. Arch and its family are already on 7.2 and will pick up 7.3 within days of the stable release. Ubuntu 26.04 LTS shipped with 7.0 and will offer newer kernels through its hardware enablement stack and later releases rather than swapping the base kernel underneath you. Fedora, true to form, launched Fedora 44 on an earlier kernel and then moved users forward to 7.0 through updates, and it will do the same dance with 7.3 once it is out. The enterprise world, RHEL 10 with AlmaLinux and Rocky, tracks the 7.0 baseline and backports selectively, so it will not adopt 7.3 wholesale. Debian stable, as always, stays comfortably behind and prioritizes predictability over novelty. If you want 7.3 on release day, a rolling distribution or a manual build is still the only guaranteed path.
Why 7.3 Matters More Than Its Number
Most kernel releases are incremental by design, and it is easy to shrug at another point release. What makes this one worth marking is the combination. The performance work is broad and real, touching the filesystems and memory paths that every workload leans on. The hardware support pushes Linux further onto machines it historically struggled with, Apple Silicon above all. And the AI question, far from being a distraction, is the kernel community doing in public what most organizations are still avoiding, which is deciding on the record how automated tools fit into serious engineering.
There is one more reason to keep an eye on it. 7.3 is a likely long term support kernel, the sort of release that gets maintained for years and becomes the stable base under countless servers, appliances and distributions. If that holds, the speedups landing now are not a temporary treat. They are the floor that a large slice of the computing world will stand on for a long time.
What to Watch Between Now and October
The release candidates run weekly until the code settles, with the stable release expected in the second half of October 2026. The things most likely to change before then are late hardware fixes and any regressions the testers surface, not the big performance work, which is already merged. If you maintain infrastructure, this is the moment to start testing 7.3 in a lab, especially if you rely on Btrfs or EXT4, because the storage gains are the kind you will actually feel. We will update this article when 7.3 ships and the final feature set is locked.
Sources and Further Reading
The performance figures and merge window overview come from Phoronix and its Btrfs and filesystem benchmarking, and from It's FOSS. The sched_ext feature complete milestone is documented in Phoronix. Torvalds and the AI in the kernel debate are reported by Tom's Hardware. Distribution status draws on UbuntuHandbook.
Published September 2026, based on the Linux 7.3 release candidates. Details may change before the final release. This is analysis and commentary, not a sponsored post. CodeHelper has no commercial relationship with the companies mentioned.