Back to Blog

Proxmox's new kernel no longer comes from Ubuntu: 7.0.14-30 vs 7.3.5-1

Inside of a rack server with the lid off: heatsinks, memory banks and the hot-swap fans in their carriers

A Proxmox forum thread from Wednesday 18 November carries one sentence that changes more than the version number it announces: "Unlike our previous kernels, which were based on the Ubuntu kernel, the 7.3 kernel is based directly on the upstream stable branch from kernel.org". This is not a release note: it is the supply chain of the kernel that boots your virtual machines, and today you can read the whole thing in two strings sitting in the repository: 7.0.14-30 and 7.3.5-1.

We run Proxmox VE clusters with Ceph in production across several datacenters, and we resell nobody's licences. So this is neither a celebration nor an alarm. It is a question that almost never shows up in a hypervisor comparison and that nonetheless decides how long you will keep receiving fixes: who builds the kernel your hypervisor boots, and for how long do they maintain it?

Two version strings, two maintenance models

We checked this today, 21 November, against the Packages file of the pve-no-subscription repository for Trixie. The default kernel metapackage, proxmox-kernel-7.0, is at version 7.0.14-30. The new one, proxmox-kernel-7.3, is at 7.3.5-1, and depends on proxmox-kernel-7.3.5-1-pve-signed | proxmox-kernel-7.3.5-1-pve.

Those strings break into two halves: the upstream point release on the left, Proxmox's own revision on the right. The repository holds thirty 7.0.14 kernel packages, from -1 to -30: a single upstream point release, repackaged thirty times. And five of 7.3 —7.3.0, 7.3.1, 7.3.3, 7.3.4 and 7.3.5— each with its own -1 suffix: five different upstream point releases, packaged once each. Same product, two opposite ways of maintaining the most critical piece of the hypervisor.

Why 7.0 carries thirty revisions: upstream declared it dead in June

The thirty revisions are not a whim. The 7.0 series on kernel.org shipped on 13 April 2026 and its final point release, 7.0.14, came out on 27 June. Seventy-five days of life and fourteen point releases. Since then, the kernel Proxmox recommends for production is a series upstream stopped maintaining five months ago, and what still arrives arrives because somebody carries it over by hand.

The changelog for the 7.0.14-30 entry, dated 16 November, says it plainly: "update sources to Ubuntu-7.0.0-44.44 plus the upstream stable changes up to v7.3.3 and v6.18.62 that Ubuntu prepared for its next update, bringing fixes for about 130 published security issues". Read it slowly: the kernel booting your production today is a 7.0.0 Ubuntu base with stable changes from v7.3.3 grafted onto it, via two intermediaries. That this tree is called 7.0.14 is, by now, a label.

And the next sentence in that same changelog is the cost of the model, written by the people who pay it: "The r8169 conversion to phylink, which comes back with them, stays reverted, as it broke the link of some RTL8125B NICs". In other words: the batch of changes once again brought a rewrite of the driver for some very common network cards, and it had to be pulled back out by hand because it killed the link on real cards. That is what maintaining a tree by grafting means, and it is where the eight faults with their own "since" we wrote about three days ago come from: regressions that did not come from upstream but from the series itself. That Proxmox's kernel came from Ubuntu's and not Debian's we already covered back in August; what changes now is where the next one comes from.

The announcement explains the reason for the change in a sentence worth reading in full, because it promises no magic: "Basing our kernel directly on the upstream stable branch shortens the path in both directions even further: upstream fixes reach our kernel with less effort on our side, and what we and our community find and fix applies directly to the code that upstream maintains". The first direction is already measurable: kernel.org released 7.3.5 on 17 November, and today, the 21st, the metapackage in the no-subscription repository is already 7.3.5-1. Four days or fewer, from the stable branch to your apt.

"Likely LTS" is not a support date

Here is the part that decides whether this move suits you. The announcement says 7.3 is "likely to become this year's upstream longterm (LTS) kernel". Likely. And that word is doing real work, because as of today there is no 7.x anywhere in the kernel.org longterm table:

Longterm series Released Projected EOL Latest point release
6.1830-11-2025Dec 20286.18.64 (17-11-2026)
6.1217-11-2024Dec 20286.12.118 (17-11-2026)
6.629-10-2023Dec 20276.6.162 (06-11-2026)
6.111-12-2022Dec 20276.1.192 (06-11-2026)
5.1531-10-2021Dec 20265.15.225 (06-11-2026)
5.1013-12-2020Dec 20265.10.274 (06-11-2026)

Look at the second-to-last row calmly, because it is what the label actually buys you: a kernel from October 2021 still getting point releases on 6 November 2026. kernel.org describes that window as one that starts at two years and gets extended depending on industry demand. Five years on, 5.15 is still being patched.

The other side kernel.org states plainly: "There are usually only a few bugfix kernel releases until next mainline kernel becomes available -- unless it is designated a longterm maintenance kernel". And the 2026 7.x series offers three consecutive examples, all checkable in the archive:

Stable series Released Final point release Lifespan
7.013-04-20267.0.14 (27-06-2026)75 days, 14 point releases
7.114-06-20267.1.13 (02-09-2026)80 days, 13 point releases
7.217-08-20267.2.15 (01-11-2026)76 days, 15 point releases

Three consecutive series dead between 75 and 80 days. That is the gap at stake: one label separates seventy-five days from five years. And 7.3 has not been given it yet.

The arithmetic that gives us pause

kernel.org publishes no calendar rule. What it does publish, answering "what is the next longterm release going to be", is this: "Longterm kernels are picked based on various factors -- major new features, popular commercial distribution needs, device manufacturer demand, maintainer workload and availability, etc. You can roughly estimate when the new longterm version will become available based on how much time has elapsed since the last longterm version was chosen". The last one was picked with 6.18, from 30 November 2025: almost twelve months ago. By that criterion, one is due. What it does not say is which.

There is a pattern you can check against the table above without interpreting anything: the six longterm series still alive are, without exception, the final mainline release of their calendar year. And the 2026 calendar, as it stands in the kernel.org archive, runs like this: 6.19 on 9 February, 7.0 on 13 April, 7.1 on 14 June, 7.2 on 17 August and 7.3 on 19 October. Sixty-three, sixty-two, sixty-four and sixty-three days between one and the next. At that cadence, 7.4 —already at rc3 on 15 November— lands in December, and then the year's last mainline would be 7.4, not 7.3.

Except for one detail that cuts the other way, and that is also in the archive: when a mainline lands at the end of October, the next one slips into January because the cycle runs across the holidays. 6.6 shipped on 30 October 2023 and 6.7 on 8 January (70 days). 5.15 on 31 October 2021 and 5.16 on 10 January (71 days). 7.3 shipped on 19 October: seventy days lands on 28 December. One week of slippage and the year's last is 7.3 again.

So we are not saying Proxmox got it wrong: it wrote "likely", which is exactly the right word, and the designation is made by the stable maintainers —Greg Kroah-Hartman and Sasha Levin— not by a blog. What we are saying is operational: "likely LTS" is not a figure you can put into a 2027 plan. If the label lands on 7.4, then 7.3 stops getting point releases about two weeks after the next one ships —that is how 7.0, 7.1 and 7.2 died— and we are back to the grafting model, just with a different number. The good news is that this resolves in weeks, not quarters.

What the announcement says, in full

The thread sells nothing, and it is worth quoting in full rather than keeping whichever half suits the argument:

«We have run this kernel on some of our test setups over the last few weeks without encountering any significant issues. The general approach of basing our kernel directly on upstream stable has been in testing for longer, as we already packaged and evaluated the upstream 7.2 kernel this way. However, for production setups, we strongly recommend either using the 7.0-based kernel or thoroughly testing on similar hardware/setups before upgrading any production nodes to 7.3.»

Both halves matter. The first says this is not an improvised experiment: they had already packaged and evaluated 7.2 the same way. The second says, in the original's own bold, that in production you should stay on 7.0 or test thoroughly on similar hardware. 7.0 remains the default kernel and 7.3 is opt-in, installed by hand with apt install proxmox-kernel-7.3. Under known issues: "None at the time of writing".

Two more warnings almost nobody quotes. The first spells out exactly what can break and why: "As this is our first kernel based directly on upstream stable, we are especially interested in anything that worked with the 7.0 kernel but no longer does with 7.3, such as missing drivers, modules, or features". An Ubuntu downstream patch or a build config option that is no longer there rarely breaks the kernel: it breaks what hangs off the kernel, which is harder to spot from the dashboard. A node that boots and a card that no longer brings up the bond is a useless node with a green dashboard — and the r8169 revert above is proof that this risk is not theoretical.

The second is purely operational and decides for you whether any of this applies: "The new 7.3-based opt-in kernel will not be made available for older release series". We confirmed it against the repository: there is not a single proxmox-kernel-7.3 package in bookworm. If you are on Proxmox VE 8, you do not have this option, and what is in front of you is not a kernel decision: it is a major-upgrade decision.

How we are handling it

On no production node, and not planned while 7.0 remains the default kernel and the estate runs fine on it. On hardware that resembles production, yes — which is what the announcement asks for and is not the same thing as a laptop. What we look at, in this order:

  1. That out-of-tree modules rebuild against the new kernel — and not "the package installs", but that they load after a reboot.
  2. The network cards, with the r8169 precedent in hand: that interface names do not change, that the bond comes up, and that the link negotiates the same speed it negotiated before.
  3. Device pass-through to guests, if the node uses it. It is the first thing to break when a build option changes.
  4. Pinning which kernel boots, using Proxmox's boot tool. An opt-in kernel installed but not pinned is a roulette wheel: the node will boot whichever one the bootloader considers newest, and that is not a decision a routine apt upgrade should be making.

If you are planning your exit from VMware, this goes in your plan

Hypervisor comparisons fight over the feature list and the licence price. Then years pass and what actually costs you work is something else: who maintains the pieces, on what horizon, and how far from the original project. In a VMware to Proxmox migration signed this quarter and executed next year, the kernel question has three parts and all three fit in one email: which kernel does the destination boot, who maintains it, and until when.

Today's answer: 7.0, an Ubuntu base, a series dead upstream since June, kept alive by Proxmox through thirty grafts. The 2027 answer could be considerably better —an upstream stable branch, with a long-term support label and four days of distance to your apt— and that is why we think this move is the right one. But that answer is not written yet, and it is not yours to decide. What is yours to decide is whether your infrastructure is built to change kernels without ceremony: a lab node with the same iron, a pinned boot entry, node-by-node updates and the ability to evacuate a host without stopping the service. With that in place, a change of kernel base is maintenance paperwork. Without it, it is an outage waiting for a date.

When you should not touch any of this

There is one clear case for installing 7.3 today: recent hardware that 7.0 does not handle properly, or a driver whose fix lives in the upstream branch and will never reach a tree called 7.0.14. If that is you, this announcement opens a route you did not have, and the recommendation to test on similar hardware is the roadmap.

For everyone else, the recommendation is boring and we stand by it: do not move. A stable estate on the default kernel, with no new hardware waiting for support, gains nothing from being first onto a branch whose long-term support label has not been handed out. What is worth doing is putting a date on the calendar —December, when 7.4 ships— and looking then at which series took the label. It is a bet that settles itself in a few weeks, and until it settles, 7.0 keeps booting your machines exactly as it did yesterday.

Sources (verified 21 Nov 2026): every English quote from the announcement —the kernel base, the "shortens the path in both directions even further", the "likely to become this year's upstream longterm (LTS) kernel", the full paragraph about test setups and the production recommendation, the sentence about this being their first kernel on upstream stable plus the feedback request, the "None at the time of writing" and the bullet about older release series— comes from the "Opt-in Linux 7.3 Kernel for Proxmox VE 9 available" thread, posted 18 Nov 2026 by t.lamprecht (Proxmox Staff). Versions 7.0.14-30 and 7.3.5-1, the Depends line, the count of thirty and five kernel packages, and the absence of proxmox-kernel-7.3 in bookworm, read from the Packages files of trixie and bookworm. The two changelog quotes —sources from Ubuntu-7.0.0-44.44 with stable changes up to v7.3.3, and the r8169 revert— from the pve-kernel changelog. The longterm table, the sentence on non-longterm stable series and both LTS selection criteria — kernel.org releases page. Latest point releases of each living series and 7.4-rc3 dated 15 Nov 2026 — releases.json. Release dates and final point releases for 6.6, 6.7, 6.19, 7.0, 7.1, 7.2 and 7.3, and for 5.15 and 5.16 — the v7.x, v6.x and v5.x archive indexes. Our own arithmetic, nobody's published figure: the 63, 62, 64 and 63-day intervals; the 75, 80 and 76-day lifespans; the 70 and 71 days of the two January precedents; the "four days or fewer" between upstream 7.3.5 and the metapackage; the pattern that the six living longterm series are each their year's final mainline; and the estimate that final 7.4 lands in December 2026. The next LTS designation has not been announced: the stable maintainers decide it.

Do you know which kernel each node of your cluster boots?

At everyWAN we run Proxmox VE with Ceph in production and we have been migrating off VMware for years. We resell nobody's licences: if the answer to that question is "we would have to check", better check before the next reboot.

Talk to everyWAN

Tags:

Share:

Subscribe to our newsletter

To receive IT stories, everyWAN news and exclusive subscriber offers, sign up to our mailing list

Minorisa de Sistemas Informaticos y Gestión S.L. © 2026
everyWAN
everyWAN