Linux 7.4-rc1 posts the largest merge window in kernel history, driven by AI-generated code
On November 1, 2026, Linus Torvalds closed a merge window of nearly 20,000 commits for Linux 7.4-rc1 — a record across 35 years — which he attributes in part to AI-triggered work. For teams shipping on Linux, that is a signal about review capacity, not a reason to rush an upgrade.
November 1, 2026. Linus Torvalds ships Linux 7.4-rc1 and closes a merge window of nearly 20,000 commits. November 1, 2026. It is the largest intake in the roughly 35-year history of the kernel — the recent average sat around 15,000 commits. November 1, 2026. Torvalds credits part of the volume to AI-triggered work, and calls its early arrival “a much healthier model.” Why it matters: the kernel’s review cadence is becoming the real bottleneck, and 7.4 is widely expected to become the next LTS.
The largest merge window in history, by the numbers
The kernel’s rhythm has not moved in decades. A major release opens a two-week merge window: subsystem maintainers submit their trees to Torvalds, who pulls them. Once the window closes, release candidates begin — usually seven or eight, one per Sunday — reserved for fixes. A large merge window therefore means a lot of code landed fast, before anyone had time to shake out the bugs.
7.4 breaks the counter. Tallies differ depending on whether merge commits are included: 19,671 commits in total, or 18,591 changesets excluding merges — either way, roughly 30% above the usual 15,000 baseline. 7.3, which held the previous record, hovered near 16,000. The gap against the long-term trend is clear.
Torvalds did not downplay it in his LKML announcement. “This is a BIG merge window,” he wrote, adding that he had left two pull requests aside to handle in rc2 rather than vacillate. The message was all about scale — and about what explains it.
What AI is doing to the development process
The cause Torvalds points to is not a new subsystem: it is AI. He was explicit about the timing of that work. “I’m actually taking it as a good sign: I’m much happier with the flood of AI triggered work going into the merge window than it showing up mainly as extra fixes during the -rc period,” he wrote. He closed with: “I think that’s a much healthier model. But we’ll see how the -rc’s look.”
The reasoning is about timing, not quality. If AI-sourced fixes arrive after rc1, they compete with the real regressions the candidates exist to catch — signal turns to noise. If they arrive inside the merge window, they pass through subsystem maintainers first, people who know the code, and reach Torvalds in organized batches. The stabilization weeks can then stay focused on actual regressions.
The pressure is measurable. In September 2026 alone, AI-generated code accounted for roughly 17% of kernel patches. Maintainers have said publicly they feel “completely overwhelmed” by waves of AI-generated bug reports and fixes. Torvalds himself concedes that AI-suggested patches are “often rewritten” by humans — a measured way of saying much of the raw output still needs cleanup. His optimism is therefore conditional: it only holds if human review stays in the loop.
The debate is already running inside the community. Maintainers have publicly floated the idea of banning AI-generated patches from the kernel’s staging area, where the least mature code lands before promotion. Torvalds, for all his loud criticism of automated slop, chose to read this volume as a good sign — but strictly on the condition that human review does not recede. The parallel with Microsoft is instructive: the same pressure drove the vendor to ship 1,461 fixes in a single Patch Tuesday, largely from bugs found by AI tools. The volume is the symptom of one shared movement, in two giant codebases that both bottleneck on the same thing: human reviewers’ time.
What is actually landing in 7.4
Behind the counters sits real content. Hardware support gets a big haul: Apple M4 (including the A18 Pro-based “MacBook Neo”, not yet ready for daily use), the start of AMD Ryzen Zen 6, HDMI 2.1 FRL, FreeSync, ALLM and VRR on AMDGPU, Google Tensor G5 and Pixel 10 in mainline, and more Snapdragon X2 laptops.
The performance numbers are flashier still. FAT32 reportedly gains up to ~265× in some cases. The early-boot allocator memblock reportedly drops from about 268 seconds to ~50 ms. Btrfs gains zstd decompression wins, EXT4 receives buffered iomap support, and SMB compression becomes toggleable via chattr. On the security side, Landlock gets composable controls. Networking lands TCP ROCCET, a CUBIC variant aimed at cellular networks.
What it means for anyone not running an rc
The calendar matters more than the features. As the last major release of the calendar year, 7.4 is the natural LTS pick — Greg Kroah-Hartman is set to take it over in early 2027. The final release is expected around December 20, 2026, or December 27 if eight candidates ship. If you are in production, stay on 7.3.1, released October 30; 7.2.15 has reached end of life.
A large intake does not break the schedule, but it loads more risk into the weeks that follow. The question is not “should we adopt 7.4” — it is whether the project’s review capacity absorbs a third more volume without letting a regression through. The coming release candidates are precisely that test.
That test has a concrete shape. If the -rc series runs long — eight or nine candidates instead of the usual seven — it will be a sign that the merged volume exceeded what the stabilization weeks could absorb cleanly. If it stays at seven and calms on schedule, it validates Torvalds’ bet that early, reviewed AI work beats late, urgent fixes. Either way, the signal will live in the cadence, not in the commit count.
Verdict
The 7.4 merge window is first and foremost a process story. Torvalds is right on the substance: an AI influx reviewed in the normal flow beats a late flood of fixes during stabilization. But the bet stays open until the -rcs speak.
If you maintain modules or distributions, watch the -rc cadence of 7.4 more than the commit count: that cadence will tell you whether human review held up under the volume. If you are in production, do not adopt a release candidate — stay on stable 7.3.x and plan your move to the 7.4 LTS once the final ships, knowing its probable LTS status makes it the long-term target. The real signal is not the kernel’s size: it is that human review, not code generation, has become the bottleneck.