# Linux 7.4 opens its merge window with 6,753 changes merged

> The Linux 7.4 merge window has accumulated 6,753 changes excluding merges, and Torvalds is considering extending it by a week due to his family vacation.
> By José Maldonado · Technology · 2026-10-24

Every few months, the Linux kernel opens a door that only lets new features through for two weeks, and whatever does not cross that threshold waits for the next cycle. The door for version 7.4 opened on Monday, October 19, 2026, a day after the release of 7.3, and four days later [6,753 changes had already been merged, not counting merges](https://lwn.net/Articles/1101211/). Most of the run remained ahead, and the main tree was still receiving entire branches of code.

That door is the merge window, the period in which Linus Torvalds merges into the main tree the pull requests that each subsystem has prepared over months. This time the routine changes for a domestic reason: [Torvalds released 7.3 on Sunday, October 18, and announced that he was starting his family vacation](https://lore.kernel.org/all/CAHk-=wgwnBm2zDdRM7u016NVHkv9KqbZW7AZ=8wZrqGyv_V+5A@mail.gmail.com/), with Monday reserved for a boat and diving, and merges pushed to the afternoons.

The interest lies less in the figure than in the pace. Anyone packaging a distribution, maintaining an out-of-tree driver or planning a server migration needs to know when the main tree moves, and a one-week difference throws off that calculation. Torvalds made clear in the same message that the extension is for him: whatever is not ready before the window opens waits for 7.5.

## The calendar matters more than the count

The window opened on Monday the 19th and, with the usual two weeks, would close on Sunday, November 1; with the extra week it would run to the 8th. Within that period Torvalds does not merge just anything: he requires that what is sent arrive tested and accepted by the subsystem maintainer before the door opens.

The previous cycle serves as a scale. Its window closed on August 30 with the first release candidate of 7.3, and the final version did not appear until October 18, seven weeks later. An extra week does not change that order of magnitude, though it does change the date on which a distribution can plan its next jump.

Each subsystem accumulates its patches in its own tree and sends them in bulk when the date arrives. Hence a delay by the one doing the merging does not slow the maintainers but jams the intake queue: the work is done and waits its turn in a list that only advances at one person's pace.

### Linus restates the rules

Here it is worth clarifying what the extension means. Torvalds himself wrote that the additional week will not make him accept pull requests that arrive outside the calendar, because what is merged must be prepared before the opening and not developed during it. His margin serves to clear the queue he himself could not attend to in the first days.

That distinction matters more than it seems at first glance: the window may last three weeks for the one doing the merging and remain closed for anyone arriving late with a half-tested driver or an unreviewed series.

The practical effect is measured in the date of rc1, the first release candidate for version 7.4. If the close moves to November 8, the entire stabilization period shifts by a week, and anyone with a freeze in sight will have to choose between waiting for 7.4 or integrating 7.3, a kernel already closed and released.

## What has been merged in the first days?

The count from Friday the 23rd covers about four and a half days of merges, close to a third of the usual calendar. The previous cycle serves to calibrate it: [the 7.3 window ended with 15,267 changes excluding merges](https://lwn.net/Articles/1089791/), the second rc1 with the most commits in kernel history after 6.7, which carried almost 3,000 from the bcachefs history.

That total was completed in stages. Some 13,000 of those 15,267 arrived after the first balance of the window, when the large pulls from each subsystem entered the main tree at once. This week's balance is a down payment, not the final portrait of the cycle.

None of those 6,753 changes is within reach of a user yet. The main tree is not packaged for production and the first version that can be seriously tested is rc1, so the figure measures the pace of the maintainers, not the quality of the result.

### More improvements for AI and servers

Among the pulls already in, the scheduler one arrived with an EEVDF update signed by Vincent Guittot, an engineer at Linaro, which in the [tests published by Phoronix](https://www.phoronix.com/news/Linux-7.4) cuts the latency of short-slice tasks, those that occupy the CPU for less time, by up to 55%. It is the same family of adjustments the engineer had already worked on for 7.3.

The same pull incorporates a mechanism that adjusts vCPUs to the time the hypervisor takes away from the virtual machine, and a preference for the two threads that share a core on NVIDIA Vera CPUs, with gains of between 3% and 6% in the same outlet's tests. The weight of these improvements is in servers and virtual machines, not in home use.

The scheduler numbers come from a specific test bench, not from a general promise. The bulk of the 6,753 changes are cleanups, bug fixes and security fixes that alter nothing visible to those who use the desktop daily.

## What changes for those who maintain a distro

No one packaging a distribution has to look at 7.4 this week. [kernel.org is still serving patches for six longterm series](https://www.kernel.org/): 6.18.58, 6.12.114 and 6.6.160 came out on October 21, and 6.1.190, 5.15.223 and 5.10.272 on the 15th. 7.2.12 appeared on the 21st, three days after 7.3.

Short-cycle distributions are the ones that do watch the main tree, and their reference now is 7.3, which [arrived with FailFS and gaming improvements for low VRAM](https://observatorioblockchain.com/en/technology/linux-7-3-failfs-scheduler-gaming-vram/), not 7.4. That kernel needs several weeks of testing before landing in a stable repository, so the window extension barely alters its roadmap for the next quarter.

The six longterm series draw a scale worth keeping in mind. They range from 5.10 to 6.18, and each carries its own revision calendar, so an administrator can go years without touching the main branch and still receive security fixes.

### Tightening the calendar

For anyone maintaining an out-of-tree driver, the date does bite. The third week does not open a second chance: if the series did not get in during the first two, the work waits for 7.5 and must be tracked in linux-next, which on October 23 published its snapshot next-20261023. That is where a long window shows, in the lag of a full cycle for new hardware.

Testing a freshly released kernel is not solved by downloading the tarball and booting, because each distribution recompiles its modules, signs the package and repeats its battery of tests with the hardware at hand. That work only begins when the kernel stops moving.

Packaging does not live on the kernel alone. Rust 1.100 will arrive on November 12 with a [minimum publication age in Cargo](https://observatorioblockchain.com/en/cybersecurity/rust-1-100-cargo-cuarentena-dependencias/), an adjustment that allows rejecting versions recently uploaded to the registry until they meet a minimum age. That kind of calendar also decides what enters a distribution and when.

## A week of margin that decides the rest of the cycle

With the window a third of the way through, the number that will go down in history is the closing one, not Friday the 23rd's. If the extension is confirmed, comparing the 7.4 total with the 15,267 that closed 7.3 will be a statistical trap: it would be three weeks of merges versus two, and the total would say more about the calendar than about the real volume of work.

What is decided is the next appointment. 7.4-rc1 will arrive on November 1 or 8, and that date drags everything else along: the length of the testing phase, the stabilization window of the distributions and the moment when a new driver stops depending on out-of-tree patches. Until Torvalds closes the door, any balance of 7.4 is provisional.
