Linux 7.4-rc1: 19,671 commits, the biggest merge window in history, and AI patches up 2.4x
🔎 A 35-year record, broken on November 1st
On November 1st, 2026, Linus Torvalds released Linux 7.4-rc1 with an unambiguous opening message: "This is a BIG merge window". Nearly 20,000 commits, compared to the usual ~15,000. In almost 35 years of kernel history, no merge window had ever reached this volume.
The exact figure: 19,671 commits between the 7.3 and 7.4-rc1 tags, a +21% increase over the previous record (16,267 for the 7.3 window). The shortlog is 24,000 lines and 1.2 MB. The source tree grows from 41,019,555 lines to nearly 42 million, according to Phoronix.
But the number stirring up the community isn't the total. It's the number of AI-assisted commits: up 2.4x in a single cycle. And Torvalds himself confirmed that he uses Claude, from Anthropic, on parts of the kernel — with three bugs discovered along the way.
Why now? Because the shift is visible. AI is no longer a curiosity in kernel development, it's infrastructure. And this window shows, figures in hand, what it produces: some very good stuff, some very mediocre stuff, and maintainers on the verge of asphyxiation.
The Essentials
- Absolute record: 19,671 commits between v7.3 and v7.4-rc1, the largest merge window in the kernel's ~35-year history (+21% vs 7.3).
- AI multiplied by 2.4: 3,008 commits tagged Assisted-by (16.2% of non-merge commits) versus 1,257 in 7.3 — and the tag was abolished in the middle of the window, on October 20, 2026.
- Torvalds uses Claude: prototyping of a file-read optimization and removal of obsolete functions, with three bugs found along the way. He calls it a "much healthier model."
- Maintainers overwhelmed: Jakub Kicinski (networking subsystem): "We are completely overwhelmed" — a third to half of the submissions to his branch are low-priority AI patches.
- Proof of value: the OverlayFS series, merged on October 31, contained an information leak, a use-after-free, and a null deref — all found through AI-assisted review.
- Timeline: seven or eight release candidates before the final version, expected in late December 2026.
Recommended tools
Want to reproduce the "agent → patch → review" workflow everyone is talking about? Here's the minimal kit, from your workstation to your test VPS.
| Tool | Main use | Price (November 2026) | Ideal for |
|---|---|---|---|
| Claude Code | Prototyping patches, cleaning up code, hunting bugs — the workflow used by Torvalds | Included in Claude Pro from $20/month (check on anthropic.com) | Kernel contributors and systems developers |
| Hostinger | VPS for compiling and booting a 7.4-rc kernel under QEMU without risking your machine | From ~$5/month (check on hostinger.com) | Testing RCs safely |
| Kimi K2.6 (Moonshot AI) | Self-hosted diff review, open weights | Self-hosted: GPU cost | Teams that don't want to send anything to the cloud |
| GLM-5 Reasoning (Z.AI) | Self-hosted alternative for patch analysis | Self-hosted: GPU cost | Constrained environments, data sovereignty |
The numbers behind an extraordinary merge window
Linux 7.4-rc1 sets the all-time record for a merge window: 19 671 commits, a +21% jump over the previous record. That's the starting point for everything else — including the AI debate.
The growth is steady and brutal. The 7.1 window had 13 963 commits, 7.2 had 14 395, and 7.3 set an initial record at 16 267. The 7.4 window crushes everything. Phoronix and XenoSpectrum report the same figures.
| Window | Commits | Verdict |
|---|---|---|
| 7.1 | 13 963 | Normal cycle |
| 7.2 | 14 395 | Normal cycle |
| 7.3 | 16 267 | Previous record |
| 7.4-rc1 | 19 671 | All-time record, +21% |
Torvalds offered two partial explanations on LKML. First, a third of the diff is nothing but register offsets and bitmasks for AMD GPUs — mechanical bulk, generated from hardware headers, which inflates the counters without saying much of anything.
Second, the LTS hypothesis: many contributors were waiting for 7.4 as the next long-term support release, and therefore pushed their work into this window rather than the previous one. The 24,000-line shortlog is the visible trace of that.
Neowin reminds us of the timeline: seven or eight release candidates before the general release. In other words, a longer-than-usual stabilization cycle to digest a monstrous diff. LWN.net tracks the window merge by merge for anyone who wants the per-subsystem detail.
Why this record is a bit misleading
My take: the number is real, but it needs to be read correctly. Strip out the AMD register dumps and the "let's wait for the LTS" effect, and you're left with a window that's still well above average — but not at +21%.
The real signal of this release isn't the volume. It's the composition of that volume, and in particular the share that AI now accounts for within it.
AI in the kernel: what the numbers actually say
About one in six non-merge commits declares AI assistance in this window — and that's a floor, not a ceiling. It's the most important figure of the release, and the most misread.
The count, established by XenoSpectrum: 3,008 commits carrying the Assisted-by tag in the 7.4 window, versus 1,257 in 7.3. A jump of 2.4x in a single cycle. In proportional terms, that represents 16.2% of non-merge commits, versus 8.2% during the 7.3 window and 19.3% during the 7.3 rc.
The catch, and it's a big one: the Assisted-by tag was abolished on October 20, 2026, in the middle of the merge window. The 3,008 commits therefore only cover part of the cycle. Actual AI usage is mechanically underestimated.
To that, add the statement from David Airlie, the graphics maintainer, at the Maintainers Summit: probably 90% of contributors use LLMs. When the maintainer of a subsystem as large as DRM says that, the question "is AI adopted?" is settled. The real question is now: how to keep it under control.
The project's answer lies in the coding-assistants.rst rule: AI agents are not allowed to add the Signed-off-by. Only a human certifies the DCO — the declaration that they have the right to submit this code. AI prepares, the human signs. Not the other way around.
Torvalds confirmed the trend in his release email: lots of fixes for problems found by AI, and more and more patches themselves suggested by AI — often rewritten afterwards by a human. It's that "often rewritten" point that the enthusiasts forget.
To apply this kind of discipline to your own repositories — structured commits, automated reviews but signed off by humans — our guide Git + AI: automating your commits and reviews details the workflow.
You can verify the scale of the phenomenon yourself on any git repository:
# Count Assisted-by tags between two kernel versions
git log --format=%B v7.3..v7.4-rc1 | grep -ci '^assisted-by:'
Torvalds and Claude: a "healthier model," according to him
Torvalds says he is "much happier" to see AI-driven work arrive during the merge window rather than as additional fixes during stabilization — and he calls it a "much healthier model." This is not just a matter of vocabulary; it's a shift in doctrine.
The context: Torvalds has been using Claude Code on the kernel since October, according to NumeriBrain. He revealed this in his announcement, with a concrete example: prototyping a file-read optimization, and the removal of obsolete functions. Along the way, three bugs were discovered.
What's interesting is the timing he values. Historically, the problem with rc periods is the late, half-baked fixes that introduce regressions just before the release. AI work that enters through the merge window — meaning early, with the full review cycle ahead of it — is, within this model, preferable.
Hence the "much healthier model": not "AI writes the kernel," but "AI-assisted work goes through the same door and the same filters as everything else." The merge window as a single funnel, rather than the end-of-cycle trickle.
This is not the first time Torvalds has shared this workflow publicly. He had already published kernel patches written by Claude Code — a symbolic moment, described in our article Linus Torvalds publie des patches du noyau Linux écrits par Claude Code : le workflow agent validé au plus haut niveau. The loop is now closed: from anecdote to regular release practice.
Let's temper that, though. Torvalds is talking about prototyping and cleanup — not writing critical drivers. Three bugs found in a file-read optimization is a nice personal validation, not a comprehensive study. The aggregate numbers, meanwhile, come from maintainers. And they are less cheerful.
"We are completely overwhelmed": maintainers under AI pressure
The problem is no longer the average quality of AI patches, it's the volume — and the maintainers of critical subsystems are cracking. The harshest words of this cycle don't come from an AI critic, but from one of its main beneficiaries.
Jakub Kicinski, the networking subsystem maintainer, wrote it bluntly: "We are completely overwhelmed, of course". According to NumeriBrain, between a third and half of the submissions to his branch are low-priority fixes, cleanups, and clarifications produced by AI agents.
His team had to negotiate a token budget to have every fix re-reviewed by several state-of-the-art models — with a pointed thank-you to Meta, which provides the resources. Think about it for a second: a Linux kernel maintainer now has to budget API calls to keep up with his patch queue. That's the new reality of the job.
The OverlayFS case: when AI review catches the heavy stuff
And AI isn't just a noise producer. It also finds serious bugs. The most striking case of the window: the OverlayFS series, merged on October 31 by Amir Goldstein, whose problems were spotted through Chris Mason's AI-assisted review.
On the menu: an information leak, a use-after-free, and a null pointer dereference — all reachable by an unprivileged user. Plus a traversal hang with remote layers and data loss with nfs_export. A good twenty commits, just for OverlayFS.
That's the dual reality of this release. AI generates a flood of marginal patches that is suffocating maintainers, and it detects flaws that purely human review would probably have let slip through. Both at the same time, in the same window. Anyone selling you a simplified version of this story hasn't read the emails.
Contributing in 2026: the new rules of the game
If you're submitting AI-assisted code to the kernel, three things have changed: the tag is no longer required, the signature must remain human, and the tolerance for noise has dropped to zero. Better to know this before sending your first patch.
First point: the Assisted-by tag was abolished on October 20, 2026. Declaring it is no longer necessary — and it wouldn't be a reliable signal anyway, since it was disappearing mid-cycle.
Second point, non-negotiable: the Signed-off-by. The coding-assistants.rst rule formally forbids it for agents. Only a human can certify the DCO. A patch signed by a bot isn't a frowned-upon patch — it's a rejected patch, and for good reason: the kernel's legal chain of certification rests on it.
Third point: target real bugs. Kicinski said exactly what saturates his subsystem — low-priority fixes, cleanups, clarifications. If your AI agent can produce forty cosmetic patches in an afternoon, the maintainer, for his part, cannot review forty of them. He is the bottleneck.
The question of code provenance has also become a security question. The Megaldon attack, which compromised more than 5,500 GitHub repositories in six hours, showed how quickly poorly verified code can spread through the supply chain. The kernel, with its growing volume of generated code, applies the same lesson in its strictest form: human review, traceable provenance, certified signature.
The ecosystem is also organizing quickly: Anthropic handed over its Agent Skills standard to the Linux Foundation, a sign that agent tooling is seeking to anchor itself within existing governance rather than alongside it.
Finally, test. A kernel RC gets booted on a disposable machine or a VPS, not on your workstation. Compile, run QEMU, verify the patch does what it claims. That's exactly what AI-assisted review managed to catch with OverlayFS — but it doesn't replace testing; it precedes it.
❌ Common Mistakes
Mistake 1: letting the AI append the Signed-off-by
The coding-assistants.rst rule forbids it: only a human can certify the DCO. An agent that signs breaks the kernel's legal certification chain — and your reputation along with it. Solution: the agent prepares the patch, you review it, you sign it. Always.
Mistake 2: flooding the lists with AI cleanups
Producing 40 cosmetic patches in a single afternoon is trivial for an agent, and impossible for a maintainer to review. Kicinski is "completely overwhelmed" precisely because of this. Solution: target real bugs, group related changes together, and read the subsystem's priorities before sending.
Mistake 3: citing "3,008 AI commits" as a measure of adoption
The Assisted-by tag was abolished on October 20, 2026, right in the middle of the window: that figure only covers part of the cycle and underestimates actual usage. Solution: present it as a floor, and cross-reference it with David Airlie's estimate — roughly 90% of contributors use LLMs.
❓ Frequently Asked Questions
When will Linux 7.4 final be released?
According to Neowin, seven or eight release candidates are planned after rc1 on November 1st. At the usual pace of one RC per week, the general release should arrive in late December 2026. The stabilization cycle will be longer than average, in order to digest the 19,671 commits.
Will Linux 7.4 be the next LTS?
That is Torvalds' explicit hypothesis: many contributors have been waiting for 7.4 as the next LTS release, which would partly explain the record size of the window. The official designation comes later; keep an eye on announcements on kernel.org.
Which AI model does Torvalds use?
Claude, from Anthropic, via Claude Code, since October 2026. He used it to prototype a file-read optimization and remove obsolete functions — discovering three bugs along the way. This is not autonomous writing, but prototyping and cleanup with human review.
Can an AI agent sign a kernel patch?
No. The coding-assistants.rst rule forbids AI agents from adding the Signed-off-by. Only a human can certify the DCO, the legal declaration that they have the right to submit the code. This is a non-negotiable red line of the project.
How many AI commits are in the 7.4 window?
3,008 commits tagged Assisted-by, i.e., 16.2% of non-merge commits, compared to 1,257 (8.2%) in 7.3 — a jump of 2.4x. But the tag was abolished mid-window: actual usage is higher, probably close to the 90% of contributors estimated by David Airlie.
✅ Conclusion
A 35-year volume record, AI adoption doubling in a single cycle, and maintainers calling for help: the 7.4-rc1 window sets the new normal for kernel development, with the numbers to prove it. To see the workflow that convinced Torvalds in detail, start with our breakdown of the patches written by Claude Code, then follow the upcoming RCs on kernel.org.
```