Sam Altman, the OpenAI CEO, said it's "quite terrible" that so much of the world's limited DRAM sits in personal computers that use it maybe 5% of the time. In his view, it would do more work in a data center.

My take? Idle RAM is the price of owning your computer. Nobody can rent it out from under you, raise its price or switch it off. That ownership is the reason this newsletter exists.

This is a short edition, though. I've been busy lately and couldn't explore much. But there's still enough here to keep your RAM busy for more than 5% of the time. Take that, Altman.

🔍 Discover AI tools

Superwhisper is now on Linux. The popular Mac dictation app now has a native Linux app for system-wide voice input. You can transcribe locally on your own CPU or GPU, or send audio to the cloud for speed. It is NOT open source software.

voiceio is for anyone who'd rather stay open source. This MIT-licensed offline dictation app streams text into whatever app you're using, on both Wayland and X11. It can also fine-tune its model on your own recordings while your computer is idle, and switches to the new model only if it scores better.

OpenKruise Agents is for running AI agent sandboxes on your own Kubernetes cluster. Interesting part is that your API keys never enter the sandbox. An egress proxy pulls them from Kubernetes Secrets and adds them to each request on its way out. It also adds L4/L7 egress rules and an allowlist for MCP tools, so a prompt-injected agent can't wander into your internal network. Overkill for a laptop. Worth a look if you self-host agents for a team.

🐧 AI in the kernel

Linux 7.4-rc1 is out, and its merge window was the biggest on record. 19,671 commits, against 16,267 for 7.3.

Torvalds credits a wave of fixes for bugs that AI found. He'd rather get them during the merge window than as late fixes in the RC stage, and called that a much healthier model. He also admits it may not be all AI. 7.4 is expected to be the next LTS kernel, and a third of the patch is AMD GPU register headers.

👀 Big Tech Watch

Now you know why Altman is eyeing your RAM.

Nikkei Asia reports that off-balance-sheet AI obligations at Alphabet, Microsoft, Amazon, Meta and Oracle grew 90% in six months, to about $3.2 trillion. All those signed data center leases that haven't started yet, plus orders for chips and servers. If they show it in the balance sheet, their shares will tank even further.

Trillions of dollars in data centers need a lot of RAM to fill them. Yours included, apparently (as per Sam Altman).

🗂 AI Jargon: KV cache

When a model generates text, it stores some working data for every token in the conversation so it doesn't have to recalculate it. That store is the KV (key-value) cache.

It grows with context length. That's why a model that runs fine in a short chat can run out of memory once you paste in a long document.

😂 Meme

Plenty more of the memes on our Instagram page.

What are you exploring on Local AI these days?

The story starts when Linus Torvalds opened the Linux 7.4 merge window more than two weeks ago when he was headed out for his annual family vacation, resulting in some temporary disruptions to his work.

He warned that his schedule would be unusual and that merging might need an extra week, but RC1 still arrived according to the usual two-week schedule.

None of that made the window any smaller. By his count, it brought in almost 20,000 commits, where 15,000 or so is the usual number, while the shortlog alone ran to 24,000 lines.

How big is it?

Phoronix went back through every merge window since Linux 3.1 in 2011 and counted the commits in each, with Linux 7.4 coming out on top at 19,671 and Linux 7.3 holding the previous high at 16,267.

That did not come out of nowhere, since every window after 7.0 has topped the one before it, with 7.4 beating 7.3 by more than 3,400 commits.

Development cycles across kernel releases show the same trend, with Linux 7.3 being the first to pass the 20,000 commits mark at 20,260 and the 7.4 window coming quite close to it.

Why?

Linus thinks AI is the reason, pointing to a surge of work set off by AI tools, including bug fixes and a growing number of patches the tools suggested, although those often end up rewritten.

He prefers receiving that work up front in the merge window over dealing with it as late fixes during the release candidate stage. He also hopes that the bigger window means AI fixes are turning into an ordinary part of development.

AI is not his only explanation. He says the expectation that 7.4 will be the next long-term support release may have played a minor role in its size, treating it as a separate driving factor.

What's included?

AMD GPU header files, which hold register definitions, account for about a third of the patch, and Linus tells people to ignore them. Drivers make up about half the patch, or roughly two thirds with the AMD headers counted.

The rest covers core areas such as networking, Rust, BPF, and memory management, along with filesystems, tooling like perf and selftests, and architecture updates.

The source tree grew along with it. By Phoronix's cloc count, the tree stood at 41,605,120 lines ahead of RC1 on November 1, up from 41,019,555 when Linux 7.3 came out.

Actual code accounts for roughly 455,000 lines of that, with the rest being blank lines and comments.

Suggested Read 📖: The latest releases of openEuler come with a heavy dose of AI agents.

My homelab runs on 5 machines with around 30 Docker containers. For a long time, the way I reached them was a bookmarks folder full of entries like http://100.101.12.4:8096.

Tailscale took care of getting me into my network from anywhere, but once I was in, I was back to remembering which box ran what on which port.

The usual fixes each have a catch. A reverse proxy with your own domain works, but it's one more thing to maintain, and every new app means touching two config files instead of one. Giving each container its own Tailscale sidecar gets you nice names, but it also turns the Machines page into a list of containers, each with its own auth key and state volume.

When Tailscale released Services, it looked like the thing I had actually wanted: names for apps, not for machines. I've been running my containers that way since, and this is a write-up of how it works, how to set it up by hand, and the two things worth knowing before you start.

What a Tailscale Service actually is

A Service is a name plus a virtual IP on your tailnet, for example jellyfin.your-tailnet.ts.net. It doesn't belong to a device. Instead, one or more of your machines advertise it, and Tailscale routes clients to them.

So your Docker host stays a single machine in the tailnet, and the apps on it become Services. Access rules can point at a Service directly, so you can let someone reach Jellyfin without giving them the whole host.

Here is what a request looks like once it's set up:

Two details in that picture matter later: tailscaled on the host terminates HTTPS and talks plain HTTP to your app, and it adds an X-Forwarded-For header. To your app, Tailscale behaves like a reverse proxy.

Setting one up by hand

Before automating anything, it's worth doing it manually once for a single container. You need three things.

1. A tagged host. Services are advertised by tagged machines, so give your Docker host a tag:

sudo tailscale up --advertise-tags=tag:server --reset

2. A policy that allows it. In the admin console under Access controls, the tags must exist, and an auto-approver saves you from clicking "Approve" for every Service:

{
  "tagOwners": {
    "tag:server": ["autogroup:admin"],
    "tag:container": ["tag:server"]
  },
  "autoApprovers": {
    "services": {
      "tag:container": ["tag:server"]
    }
  }
}

3. The Service itself. On the Services page of the admin console, choose Advertise, then Define a Service, name it jellyfin and give it the tag:container tag. Then, on the host, point it at the app:

tailscale serve --service=svc:jellyfin --https=443 127.0.0.1:8096

This assumes Jellyfin publishes port 8096 on the host. If HTTPS certificates are enabled for your tailnet (DNS page in the admin console), https://jellyfin.your-tailnet.ts.net now works from every device that runs Tailscale.

For one app, that's fine. For 30, it gets old quickly: one definition in the console and one serve command per app, plus published ports again. And nothing updates by itself when you add, remove or move a container.

Why I ended up scripting it

The part that bothered me most was that the configuration lived in two places. The app was defined in a Compose file, and how to reach it was defined somewhere in the admin console and in tailscaled's state. Remove a container and its Service just stays there.

Traefik solved the same problem for reverse proxies years ago with Docker labels, and I didn't find anything like that for Tailscale Services. So I wrote a small Go program that does it, DockTail. It runs as one container per host, watches Docker events, and turns labels like these into the steps above:

services:
  docktail:
    image: ghcr.io/marvinvr/docktail:latest
    restart: unless-stopped
    volumes:
      - /var/run/docker.sock:/var/run/docker.sock:ro
      - /var/run/tailscale:/var/run/tailscale
    environment:
      - TAILSCALE_OAUTH_CLIENT_ID=${TAILSCALE_OAUTH_CLIENT_ID}
      - TAILSCALE_OAUTH_CLIENT_SECRET=${TAILSCALE_OAUTH_CLIENT_SECRET}

  jellyfin:
    image: jellyfin/jellyfin:latest
    restart: unless-stopped
    volumes:
      - ./config:/config
      - ./cache:/cache
      - /path/to/media:/media:ro
    labels:
      - "docktail.service.enable=true"
      - "docktail.service.name=jellyfin"
      - "docktail.service.port=8096"
      - "docktail.service.service-port=443"

With an OAuth client (Services: Write, Devices Core: Write, scoped to tag:server) it also creates the Service definition, so the manual console step goes away. Note that Jellyfin has no ports: section: the traffic goes straight to the container's IP on its Docker network.

That's also why this needs something that watches Docker. If you skip published ports and point serve at a container IP by hand, it breaks the next time the container is recreated:

DockTail is open source (AGPL-3.0). You don't need it to use Services, though, and everything in the next section applies either way.

Two things worth knowing

If a Service doesn't show up, it's almost always the tag. The host needs to be tagged (check tailscale status), and if you haven't set up an auto-approver, the Service waits for you on the Services page.

Apps that check proxy headers may complain at first, because Tailscale adds an X-Forwarded-For header. Home Assistant answers with "400: Bad Request" until you trust the proxy in configuration.yaml:

http:
  use_x_forwarded_for: true
  trusted_proxies:
    - 127.0.0.1

On a Docker bridge network, use the address Home Assistant names in its log instead of 127.0.0.1. Jellyfin has the same setting under Dashboard → Networking → Known proxies, but it works fine without it.

Was it worth switching?

For me, yes. What I like most is that the network config now sits right next to the app in the same Compose file. Adding an app means adding labels, removing it takes its Service with it, and my Machines page shows machines again.

I'd still use a reverse proxy for anything public with my own domain. Tailscale Funnel is great for a webhook or sharing one thing with a friend, but it's not a replacement for that. And if you run a handful of containers that rarely change, the manual tailscale serve route above is perfectly fine and needs no extra tool.

If you want to try the label approach, the code and docs (including complete guides for Jellyfin and Home Assistant) are at github.com/marvinvr/docktail and docktail.org/docs.

And if you've solved this differently, I'd honestly like to hear how. Leave a comment, I'll read all of them.

Building a Fedora release means tracking an enormous amount of formal change proposals planned for each release cycle on a public Change Set page, where every entry carries an owner and a status.

These proposals undergo community review and require approval from the Fedora Engineering Steering Committee (FESCo), either in batches or individually, depending on their scope.

But approval only means a change is accepted for the planned release, not that it is guaranteed to ship. The beta for Fedora 45 came out last month, and several proposals that were meant for the stable release were pushed to Fedora 46.

What we have now is a release that was initially set for October 20, but was pushed back to October 27, and then again, to today, November 3, because there were some late-stage teething issues with the release.

Better Kernel Panic Handling

If you have ever hit a kernel panic on Fedora, you might have spotted a QR code on the screen. The DRM Panic component has shown it by default since Fedora 42, but scanning it only got you a raw trace that not many people can read.

Fedora 45 changes that with a web frontend for DRM Panic. Scanning the QR code with your phone now takes you to a Fedora-hosted page that walks you through what went wrong.

From there, you can start a Bugzilla report with your system details and the trace already filled in.

Inclusion of kmscon

Similarly, if you have pressed Ctrl+Alt+F3 and landed on a full-screen text login, you have seen the Linux virtual console. That screen is drawn by the kernel's fbcon, and Fedora 45 swaps it for a userspace program called kmscon.

The switch brings back scrolling, adds basic mouse support, and a crash no longer results in a kernel panic. You also get resizable fonts, better Unicode support, and the ability to use the same keyboard layout as your desktop.

Don't worry about things breaking. If kmscon fails to launch, Fedora falls back to the old getty and fbcon setup automatically, and you can switch back yourself if you prefer.

Linux 7.2

Fedora 45 ships with Linux 7.2, and a few of its changes are easy to appreciate without digging through the changelog. Radeon users get initial HDMI 2.1 support years after the HDMI Forum blocked it from the open source driver, and the new ISP4 driver completes webcam support on the HP ZBook Ultra G1a and other upcoming Ryzen laptops.

Existing laptop users get more hardware support too. The ASUS ROG Strix G16, Zephyrus G14, and TUF Gaming F16 are now supported, Lenovo laptops can limit their battery charge, and Microsoft's 12-inch Surface Pro gets a platform driver.

Though, I must say, it's a bit weird to see Fedora skip out on Linux 7.3, given that Ubuntu 26.10 got it in before the kernel was even officially released. Perhaps the release delay had something to do with the switch, but that's just speculation on my part.

Desktop Environment Upgrades

Fedora Workstation 45 comes with GNOME 51, and its handful of practical upgrades include offline map downloads in the Maps app, a new fingerprint manager in the Users panel of Settings, and monitor brightness that is saved and restored across sessions.

The Files app picks up rename and trash options in the file previewer, and Mutter drops support for legacy NVIDIA drivers.

Fedora KDE Plasma Desktop 45 comes with Plasma 6.7, and the feature to talk about is per-screen virtual desktops. This is a useful thing for multi-monitor setups where your laptop display and external monitor no longer have to be limited to the same set of virtual displays.

Additionally, the classic Oxygen and Air themes from the KDE 4 days are back as full Global Themes, and if you have separate day and night themes set up, a new toggle in the Brightness & Color quick settings lets you switch between them.

Podman 6

Podman is a container tool that lets you run and manage containers and pods using Docker-style commands without needing to run a daemon in the background.

With this Fedora release, Podman moves up to the 6.1.x series, and most of the significant changes come from the major 6.0 release it builds on.

That release retired slirp4netns, CNI, cgroups v1, and iptables in favor of Pasta, Netavark, cgroups v2, and nftables, and it migrates an old BoltDB database to SQLite automatically.

Beyond the removals, network isolation is now on by default, podman volume prune only clears anonymous volumes that aren't in use, and Quadlet files live in their own subdirectories.

Podman 6.1 then adds smaller conveniences like podman volume rename, podman machine restart, and retry options for podman manifest push.

Miscellaneous Changes

Here are a few additional improvements that didn't make it into the sections earlier:

• A new gold-colored, gears-themed default wallpaper.
• NetworkManager now has support for IPv6-mostly networks enabled by default.
• Stratis storage can now be used during installation, with support in both Kickstart files and the Anaconda WebUI's manual partitioning.
• Fedora Atomic Desktops, such as Silverblue and Kinoite, switch from the old GTK-based Anaconda installer to the new WebUI one.
• For confidential computing on UEFI systems, there's a new stripped-down GRUB package that can quickly boot Unified Kernel Images while the regular GRUB stays in place.
• A slow move away from GnuPG with a new RPM macro that uses Sequoia by default and brings post-quantum cryptography support.

📥 Download Fedora 45

You can get the stable release from the official website. Pick the image that suits your use case, and expect there to be some minor breakage that could affect your machine.

Personally, I have never owned an Apple device, mostly due to its pricey nature and how the brand loves to lock people in their walled garden. But that's just me ranting there for a bit.

There are many who do like what Apple has to offer and are willing to see what additional value they can get out of that hardware by breaking free from macOS.

This list covers six such options that ought to awaken the tinkerer inside you. 🧑‍🔧

1. Fedora Asahi Remix

Fedora Asahi Remix is Fedora Linux with the Apple Silicon work from the Asahi Linux project built in. KDE Plasma on Wayland is the default desktop, and GNOME, Minimal, and Server images are there if you want something different.

On M1 and M2 Macs, almost everything works, from the webcam and microphones to Wi-Fi and Bluetooth. The graphics drivers also pass the OpenGL 4.6 and Vulkan 1.4 conformance tests.

M3 owners can try it too, but support is quite new. Sleep does not work, HDMI on MacBooks is unavailable, and 3D graphics don't run well yet. The M3 Ultra Mac Studio is not supported at all.

Installation begins in macOS. Run the following command in the Terminal app and follow the prompts to continue with the installation of Fedora Asahi Remix on your Mac:

curl -L https://alx.sh/ | sh

M3 owners are recommended to run dnf upgrade --refresh after the install finishes.

2. Ubuntu Asahi

If Ubuntu is already familiar to you, then this community project has ported it to M1 and M2 Macs by carrying over the Apple Silicon support from Asahi Linux.

The installer comes in four flavors, covering 26.04 and 26.10 for Ubuntu Desktop and Server. The goal is to provide the same feature set as Fedora Asahi Remix, while new hardware support only arrives after the Fedora base ships it.

As a result, you can expect system updates to be behind by a few weeks.

Like earlier, you get the installer as a script and run it in Terminal. It shrinks the existing macOS partition and lets you pick a release and build, leaving your macOS installation intact:

curl -sL https://ubuntuasahi.org/install | sh

3. Debian Bananas

Debian doesn't officially run on Apple Silicon yet, but the Bananas Team is working to change that. It packages the missing pieces and ships them from an unofficial archive of its own.

Official support is stuck on the kernel. The Asahi Linux patches it needs aren't upstream yet. Most other pieces, like the m1n1 bootloader and the audio stack, have been in Debian since Trixie.

The team's installer offers GNOME, KDE, or a console-only setup:

 curl -L https://bananas.debian.net/install | sh

The install page contains the necessary steps.

4. NixOS on Apple Silicon

Write the whole system down in a Nix config file, rebuild, and the Mac matches it. That's the NixOS way, and this community port brings it to Apple Silicon. The maintainers keep versions close to Fedora Asahi Remix and turn away changes that stray far from it.

Hardware support follows whatever Asahi Linux supports, and setting it up is quite a hands-on affair. You’ll prep a UEFI environment from macOS with the Asahi Linux installer, boot a NixOS installer from a USB drive, and finish the job from the command line.

5. Omarchy M

Omarchy M is a dedicated team rather than a separate distro that has been put together to handle Omarchy support on Macs. It's bringing the Arch-based, Hyprland desktop to Apple Silicon while also giving some attention to older Macs.

The first release targets M1 and M2 Macs, along with their Pro and Max variants. On the roadmap are a native installer, full GPU support, USB-C external monitors, Touch ID, MLX, and disk encryption.

Support enablement work for newer Macs will take place alongside this too!

The native installer is a community-built app by team member Marcelo Alcantara, and it sets up Omarchy on a Mac machine without needing a USB stick.

6. Gravity Linux

Gravity Linux is built to go away. It's a Fedora remix with KDE Plasma on Wayland, and the plan is to retire it once the drivers land upstream and any mainstream distro can run on supported Mac devices.

It started as a fork of Asahi Linux over a difference in policy on LLM use. Everything it offers comes from clean-room reverse engineering.

The first alpha release runs on the M4 Mac mini and is meant for developers. The display controller and GPU work, but suspend, USB-C display output, Thunderbolt, and USB4 don't; shutdown and reboot behavior is flaky.

You can install it on macOS with the command below, but do plan a full reinstall when the beta lands, as there's no established upgrade path.

curl -fsSL https://install.gravitylinux.org | sh

I will be adding more to this list as I come across reliable new projects. Meanwhile, if you know some, please mention it in the comments.

Adobe charges $60 a month. Microsoft 365 keeps raising its price. Salesforce will take $150 per user without blinking. These are not tools you own. They are subscriptions that expire the moment you stop paying, often taking your files and workflows with them.

The Puter team thinks that is worth fixing. Remember Puter? The folks behind the PuterOS that gives you a workable operating system in a web browser.

Their new project, called everything, is a single GitHub repository with one specific goal and that is to rebuild the world's software from scratch, free and open source.

This is what they are aiming for. Over 50 web-based applications spanning nine categories, including a photo studio, full office suite, CRM, CAD tools, video editor, and circuit design app. Every one of them is open source under the MIT-license. And they can be self-hosted, too.

Basically, it targets all the major Software as a Service (SaaS) companies.

Tapping Into Everything

The applications are all listed on the official GitHub page, they're all in a table, and very helpfully, the cell next to the application name states what popular software it is an alternative to. For example, "Notes" is an alternative to Microsoft OneNote, Evernote, Apple Notes. It is all put in well organized categories.

The apps are all hosted online, and can be opened by pressing the "Try it" link. They're completely usable, however, and in case you have a slow system that can't handle some heavy duty tasks such as video editing, this is genuinely a better option. Nevertheless, you have to remember that you'll be uploading your data to this service, and if that is a privacy concern then you may have to reconsider.

Local Hosting

On the other hand, you can also host the applications locally, the instructions for which are given on the GitHub page, again.

For example, here I am running the Video Studio locally:

cd everything/src/video-studio/
npm install
npm run dev

This returns a link you can open up in your browser and use the application.

Storage

Background

HeyPuter launched Puter in 2024 as a self-hostable internet computer: a full desktop environment running in a browser, with cloud storage, apps, and developer tools baked in. It was well received by the community.

With Everything, rather than building one great alternative to one piece of software, the project takes on the entire proprietary stack at once.

The logic is that the vendor lock-in does not come from a single app. It comes from an ecosystem. Adobe keeps you because Lightroom talks to Premiere talks to Acrobat. Microsoft keeps you because Word talks to Excel talks to Teams. Breaking one link does not free you. Everything is trying to break all of them.

The MIT license matters here. Unlike the GPL, MIT places almost no restrictions on use. You can take these apps, modify them, and run them anywhere without obligation. For businesses and developers worried about compliance, that removes the last excuse to keep paying for proprietary software.

Experiencing the Everything apps

I've tested some of the apps to check the quality of the production, and pretty much to see if the job gets done.

PDF Editor

The PDF editor has great options, adding texts or links, images, stitching PDFs together and so on.

You can even draw diagrams to add to the PDF in the app itself. You can also add sticky notes, highlight text etc. These are quite basic features on a PDF editing menu, but they work seamlessly here, which is great to see.

Since the software is in early stages, there are some bugs. Like when I tried editing the preexisting text in a PDF file, it was not entirely accurate. The text doesn't always get detected in proper blocks, and the formatting gets all over the place as a result of that.

Video Studio

I tried some of the features, such as the AI caption generation, and the film grain effect, with transitions in between clips and it worked quite well.

It also generated captions and allowed me to edit them if a word was incorrect.

The export menu showed options to separately export the audio in a WAV file and captions in an SRT file, which is great.

I had an issue with one of my video clips where it said my browser couldn't decode the video file, but then there was no further information on how to fix it, which essentially rendered the video file uneditable.

Music Studio

Being used to Ableton myself, the interface was different here for me. Don't get me wrong, the app itself is pretty intuitive to navigate through, the major actions are all laid out in a great way.

It has all the effects you might need, such as compression, reverb, delay, parametric EQ, stereo and so on. They work quite well, and the options for each are ample.

Nothing too technical, but just as much as you need.

Even the export showed options to export the stems as well as MIDI files separately, which is a great feature to have. Out of all the three that I tested, the music studio worked most to my satisfaction.

Conclusion

There is a new trend in the developer world these days to reverse engineer everything with AI. Puter's new project is part of the same trend.

The apps on The Everything Projects have their ups and downs but on the larger scale, they work incredibly well for the most part, giving features that almost match the professional level. At least for beginning professionals everywhere, who cannot afford expensive software, these apps are a great way to get things done.

It requires an account to access the AI features and to save the progress on projects unless you have an account, where you will probably need a Puter subscription because 25MB is too small a storage to actually get anything done.

That actually is understandable to avoid the service abuse but the 25 MB could have been 250 MB at least. In any case, self-hosting option is always there.

What do you think about this wave of reverse engineering apps and open sourcing them? Let us know in the comments. Cheers!

govdigital eG has awarded IONOS a four-year framework contract under KoDaS, the abbreviation for a collaboration and file storage system. The service delivered under it is built on Nextcloud and sold as Nextcloud Workspace.

It is meant for schools and school authorities in the German states of Berlin, Hamburg, Bremen, North Rhine-Westphalia, Rhineland-Palatinate, and Hesse. Other German states can join later, and schools are not mandated to implement it.

How the deal works

IONOS is the main contractor. It is tasked with delivering and operating the service, running the cloud infrastructure located in German data centers, and building it all on Nextcloud's open source software.

The tender notice for KoDaS sets a maximum value of 67.5 million euros and a maximum term of 48 months from the award. Nextcloud's announcement names the six states and their schools as the main parties who can order, leaving room for other states to be added later.

The announcement does not say what the contract will actually be worth. 📝

The notice calls for procuring software available under an OSI-approved open source license, citing digital sovereignty, lasting availability, and the ability to switch providers.

Among its asks for bidders are ISO 27001 certification, at least 10 million euros in turnover, liability insurance of at least 5 million euros, and a reference project with at least 10,000 active users and fully separated tenants.

For the features, the notice calls for central file hosting and sharing, a browser-based office suite for working on documents together, calendar and contact management, and an integrated email client.

Pricing depends on how many people actively use the service, and it is split into five tiers that start at the up to 50,000 users mark, ending above 500,000. No school authority has to commit to a set volume up front, so each can begin with its current users and grow from there.

Nextcloud's main argument is that public institutions should have alternatives to proprietary platforms so they keep control of their data. The company also believes students who start out using open source tools are less likely to reach for big-vendor products afterward.

We have seen this already

Other German institutions are also reworking their software. Last year, we heard that almost 80% of Schleswig-Holstein's state government workplaces had switched away from Microsoft.

Mecklenburg-Vorpommern is replacing SharePoint with a self-hosted Nextcloud that around 5,000 employees use today, and the state plans to bring that to 50,000 across the public sector.

Then there's the federal Deutschland-Stack, published by the Federal Ministry for Digital Transformation and Government Modernization, which covers all levels of public administration and permits use of only ODF and PDF/UA for documents.

I anticipate more such moves across Europe as governments realize that keeping sensitive data with providers that fall under foreign law can expose it to access they have no say over, irrespective of where the servers are located.

The ongoing US midterms could also shift the conversation depending on who comes out ahead.

Suggested Read 📖: France's absurd per-gigabyte tax could cost VLC over €1 million a year.

Yet another new Linux distribution has landed on Arch Linux as its foundation of choice. This one is called Pond, and it is Butterfly Interaction Inc.'s vision of how a desktop should look and behave.

For its desktop interface, it uses Niri, a very popular Wayland compositor that is known for its scrollable tiling, where app windows sit in columns along an endless strip that gradually grows as new items are added.

Pond is labeled an Early Preview, and this First Look piece covers what it offers, my time using it, and the ways you can install it.

Pond: The details ✨

Butterfly says Pond mixes new ideas with desktop habits people already know, and the goal is for a computer to feel more natural to use.

It builds that around four ideas. Windows are arranged spatially and can be rearranged, components come in different sizes to fit the space available, useful information stays in one uniform place, and content only shows up when it is relevant.

They also point out that each component was planned with the whole desktop in mind, from its proportions to where it sits and how it behaves.

As I pointed out earlier, Pond is a rolling-release distro based on Arch Linux with Niri as its compositor. Package duties are handled by a package manager that picks a default source for each app across Pacman, Nix, and Flatpak. More on that later.

Getting it to run

I tested Pond inside a QEMU virtual machine with 4 GB of RAM and a 50 GB disk, making a few tweaks to the machine settings before I could get it to boot.

Right away, the installer caught my eye. It follows a very different approach than what most distros ship with, featuring a simplified one that takes users through the steps quickly.

It asks for your keyboard, language, and timezone first, then the disk, and finally a username, password, and computer name. I really liked how entering passwords in the text field showed me some Pond-themed icons instead of the usual dots or asterisks.

The login screen follows a different approach too, with a cool-looking square and line-shaped password entry box that filled up the empty small squares as I typed my password.

The desktop

Log in, and you will see that the navigation on Pond is keyboard-focused, with an onboarding flow explaining which keybindings are must-knows for getting around Pond's interface and any installed apps.

The desktop itself features a sidebar menu that houses the clock and calendar, the currently running apps inside the minimap (the squarish-looking box), and notifications from the system, web browsers, or the package manager.

A small dock-style menu sits below it, housing quick access to the network and volume controls, as well as the settings page and the hide sidebar button.

Then there's the Vicinae-based launcher, which will feel familiar if you have used launchers on any Hyprland-equipped distro like CachyOS and Omarchy.

Installing apps

Installing new apps is easy. You open the launcher using Super+Space and go into "Install apps" to see a long list of apps that work on Pond. Select one, authenticate the action with your account password, and wait for it to finish installing.

While that happens, a banner shows up in the sidebar under notifications, showing you the progress of installation. To test this out, I installed Firefox, and it ran well. Pond served me the latest build at the time of writing, Firefox 159, and it ran without any hiccups.

Pond's package manager handles those installs, and its GitHub repo shows that most of the catalog is Flatpak apps, with the rest coming from pacman and the AUR.

Those pacman and AUR apps are installed through paru, and updates run paru and Flatpak together. Nix, while present, plays a smaller part, mostly being used to facilitate GPU support.

You also have the option to install package managers like yay and npm.

Tile away!

Window tiling is the best part of Pond. I could move around app windows using Super + Left or Right, switch workspaces using key combinations like Super + Up or Down, and change window sizes using Super + Plus or Minus.

The minimap below the calendar shows each workspace as a row, with the apps inside it as icons. I could even drag and drop app windows inside it to move them across workspaces.

Verdict 📝

The distro follows a very different design language than other Linux distros, relying on colorful shapes and palettes while offering a known, reliable base to work off of.

It is still an early preview, but if development continues and people take a liking to it, I don't see why it couldn't be counted among the top-rated new blood Linux distros alongside the likes of Bazzite, Omarchy, and Nobara.

Trying Pond 📥

Pond lists two installation routes on its install page, and downloading either one means agreeing to Butterfly's Terms of Use and Privacy Policy. If you are someone who doesn't like doing so, I suggest going through the fine print before committing to an install.

The first method would have you download an ISO of about 1.5 GB. Then you either flash it onto a USB stick and boot, or plop it into GNOME Boxes or Virtual Machine Manager to run it in a virtual machine.

The second method skips the flashing, letting you run an installer on Windows, where you pick a drive, set aside space for Pond, and set up dual boot.

Butterfly says the installer carries over supported Windows settings and handles common dual-boot problems such as clock mismatches and NTFS drives that block shrinking.

You will find the source code for Pond's various components on GitHub.

Okay, killed by Ubuntu is an exaggeration, as Ubuntu is not the only Linux distro doing this.

Still, the big news is that Linux Mint is bringing hibernation back.

In a recent blog post, the Mint team previewed a new Sleep Modes page in the System Administration tool. It's meant to make enabling hibernation and setting it up correctly easy, without you editing config files by hand. The feature is planned for the upcoming Mint 27 release coming in December.

Mint hasn't magically fixed the problems that made Ubuntu drop hibernation. What it does is give you the choice, along with two hybrid modes that sit between suspend and hibernation.

Hibernation vs suspend: What's the difference?

Both features are about saving your session, which simply means everything you have open like your apps, windows and documents.

When you suspend (also called sleep), your computer keeps your session in RAM and powers down almost everything else. Waking up takes a couple of seconds. The catch is that RAM needs power to hold that data, so a suspended laptop keeps draining its battery.

If the battery dies while suspended, your session is lost, along with any unsaved work.

Hibernation copies your session from RAM to your disk and then powers off the machine completely. It uses no power at all. When you turn the computer back on, the session loads back from disk and your apps and windows are exactly where you left them.

In short, suspend is fast but needs power. Hibernation is slow but survives a dead battery.

Why Ubuntu and other distros dropped hibernation

Ubuntu disabled hibernation by default back in Ubuntu 12.04, I think. The reason was that it was unreliable slow, and that having two suspend modes confused people.

Every Ubuntu-based distro, including Linux Mint, inherited that decision. Actually, it's not just Ubuntu. Many other Linux distros also do not offer hibernation option anymore.

Speed is indeed a problem. Writing gigabytes of RAM to disk and reading it back can take longer than a full reboot, especially on older hard drives.

Hibernation also needs a lot of swap. Your swap space has to be at least as large as the RAM in use so the whole session fits. Ubuntu ships a small swap file by default, and Fedora uses zram, which lives in RAM and is useless for hibernation.

Then there's Secure Boot. With Secure Boot enabled, the Linux kernel runs in lockdown mode, and lockdown blocks hibernation because the saved image on disk isn't signed and could be tampered with.

Since most new computers ship with Secure Boot on, hibernation simply doesn't work for a large chunk of users. This is one of the main reasons Fedora Workstation doesn't offer hibernation out of the box either.

The three sleep modes in Linux Mint

The new Sleep Modes page gives you three options. You pick the one that suits how you use your computer. You can enable all of them and choose which one to use from the power menu (at least that's what it looks to me).

Hibernation

This is the classic mode. Your session is saved to disk and the computer shuts down completely, using no power. It's slower than suspend and can take longer than a reboot to hibernate and resume.

Hybrid Sleep

Hybrid Sleep suspends your computer and saves the session to disk at the same time. Most of the time, you get the quick wake-up of suspend. If the battery runs out, you'll need to plug in the charger, but your session is still there to resume from disk.

Suspend-then-Hibernate

This one suspends first and switches to hibernation only when the battery level gets critical. I think this is the most practical option for laptop users. You get fast resume in normal use and a safety net when the battery is about to die.

What Mint's new sleep mode doesn't fix

None of this removes the old limitations. You still need plenty of swap, and the tool flags a lack of swap as a problem. If Secure Boot is enabled, the tool lists it under Detected Problems. The good thing is that the tool writes its settings to its own systemd file, so you can turn these modes on and off without messing up your system.

I like how Mint takes a subtle approaches to "new features". Not every new feature needs to have AI in it. This is not a shiny new change but it's one of those changes that everyday Linux users are likely to use. Am I right?

At least that's what VLC thinks when protesting the new French law.

France's National Assembly has voted in favour of a new "internet tax" that would charge companies for every gigabyte of data they send to users in France.

The idea is to make Netflix, Google, Meta and the rest of Big Tech pay more.

That sounds like a dream come true for many of us who prefer taking ownership of data in our own hands but unfortunately things are not that simple.

The way this law is written, it doesn't stop at Big Tech. Even VLC, a free and open source media player maintained by a non-profit, could end up paying this new tax.

What the new French internet tax says

The measure is amendment I-1358 to France's 2027 finance bill, tabled by Laurent Lhardit. It creates a tax on "producers of digital data" at €0.055 (5.5 euro cents) for every gigabyte delivered to subscribers of French internet providers.

Internet providers would measure how much data each sender pushes to their customers, with the collection supervised by telecom regulator Arcep and the tax administration DGFiP. Only the volume is counted, not the content or where it goes.

The amount is capped at 15% of the company's revenue in France or 2% of its total revenue, as per the amendment. Note that it's 15% of French revenue, not of total network traffic. And the text joins the two caps with an "or" without saying which one applies when they differ.

How much money are we talking about?

The amendment cites an Arcep estimate of 200 billion GB a year flowing to French internet subscribers. It expects the tax to bring in €10 billion a year. It also names Google, Meta (Facebook, Instagram), Amazon, Netflix and Akamai (cloud server provider) as the senders of about half of that traffic.

You'd think the money goes to the telecom operators who build and maintain the networks. That's the "fair share" argument European telcos have pushed for years.

But that's not the case. The revenue goes straight into the state budget, and the amendment's own justification talks about tax fairness between local businesses and digital multinationals, plus the ecological benefit of compressing data.

Why VLC is angry

A gigabyte of open source software counts the same as a gigabyte of Netflix. VideoLAN, the French non-profit behind VLC, posted on X that the tax would cost it over €1 million a year.

Lhardit replied that the cap would protect VideoLAN. He also said he didn't see how the tax could be passed on to users of Facebook, Instagram or TikTok since those services are free.

VideoLAN didn't take that well:

You have no idea what you are talking about. What you said about Facebook and Insta and other social networks are not targeted because they are free, means you don't understand the basics of economy. You don't understand the difference between Revenue and Profit.

VideoLAN (@videolan), October 30, 2026

The cap may limit what a low-revenue non-profit owes. But there is no explicit exemption for associations, free software or free content anywhere in the text. VideoLAN's €1 million is the possible figure before the cap, and right now nobody can tell the exact figure for certain because of all the ambiguity.

VLC isn't alone. OVHcloud founder Octave Klaba put his company's gross exposure at around €70 million a year, and Free's (Teleccom) Xavier Niel called the authors "ignares incompétents", roughly "incompetent ignoramuses". Those are French companies, by the way. Not exactly the US giants this internet tax bill was meant to tame.

At the end, the consumer pays the bill

See, a "war on Big Tech" sounds great in a headline, and European governments are fond of laws that squeeze extra money out of American tech companies. But I don't buy the idea that Netflix will pay them out of their pocket.

Netflix will do what it always does and raise subscription prices. And Facebook is "free" only because advertisers pay for it, which is exactly the revenue versus profit point VideoLAN was making.

So the government gets extra money, Big Tech protects its margins, and you end up paying. Meanwhile, non-profit projects like VLC, which never had a price to raise in the first place, are left wondering how big their bill will be.

Let's just hope that this doesn't take the form of law. Because this will set a precedence. I wouldn't be surprised to see other European countries try their own version. No government will say no to extra tax revenue.

Like GDPR, this has all the ingredients of another mess. If a law is going to tax "internet companies", it needs to say in black and white which ones. Right now, it doesn't.

What do you think? A fair way to make Big Tech pay, or a blunt instrument that hits everyone else? Let me know in the comments.

Remember Wubi from the 2000s? For people who were afraid of going into BIOS and booting from CD/DVD/USB, Wubi provided an easy way to install Ubuntu Linux straight from Windows.

This concept has made a comeback in the form of Libertix although the core idea remains the same: install Linux in dual boot from inside the Windows.

Libertix is a Windows application. You use the .exe to install the tool and run it. The tool analyzes your system and if the test passes, it lets you resize the disk to install Linux Mint or Zorin on your system. Those are the two distros supported at the moment.

It configures a dual-boot menu for both BIOS/MBR and UEFI/GPT machines. This way, you do not need to create a live USB.

Libertix: Install Linux from Windows

Libertix is an open source software developed by Ekimia, a French company. From the looks of it, they sell Linux preinstalled computers in France. They launched a €3,000 campaign in June 2026 to pay a full-time contributor over the summer to work on this project.

Here's how it works.

Before touching anything, Libertix runs a compatibility test. It looks at your firmware type (BIOS or UEFI), your disk layout, BitLocker status and how much space can be taken from the Windows partition. If something doesn't fit, it won't proceed.

If everything checks out, it lets you shrink the Windows partition and then it downloads the chosen distro ISO. Checksum is verified before downloading the ISO.

Libertix then stages the files on disk and tells your computer to boot into its own live environment just once. On UEFI systems, that's done through the BootNext firmware variable. On older BIOS systems, it uses GRUB4DOS and a temporary Windows boot entry.

After the reboot, the Libertix live system checks the plan and hardware again, creates an ext4 partition and installs Linux on it.

The final setup gives you a customized, beautiful-looking GRUB menu with entries for Linux, Windows, shutdown and advanced options.

If you enabled file sharing, your Windows folders show up in Linux. Your Linux files are also accessible from Windows, but in read-only mode, which makes sense as it avoids filesystem corruption.

I like that there's a recovery path built in. Libertix tracks every stage it completes. If something fails or you cancel midway, it undoes only the steps that were done and restores the Windows boot.

Worth a try?

I wanted to give Libertix a try and show the installer in action. But for that, I would need a functional Windows system. And I don't have one right now. I'll see if I can arrange one and showcase this tool's functioning to you.

Till then, if you feel like, you can give it a try on a spare system because this is in early stages of development. If you are not comfortable with partitions and stuff, don't get over excited and mess up your system.

Still, overall, this is a nice tool to exist and I hope it supports more distros in the future. For many Linux beginners, creating the live USB and dual booting from it seems intimidating. If this tool helps lower that barrier, I am rooting for it.

If you want to explore the project, look at its architecture and technical documentation; head over to its GitHub repo.

Collabora has published its own account of the Danish ministry pilot that earlier coverage, including ours, had described as Denmark moving off Microsoft's office software.

The ministry's own launch release had described it as a test by a group of employees in one department. Back then, the minister handling this, Caroline Stage Olsen, had said the tech giants were not being tossed out the door overnight.

A volunteer test, not a switch

The pilot began in June 2025, when volunteers in the Danish digitalization ministry's own department started writing documents in Collabora instead of Microsoft's Word, Excel, and PowerPoint editors.

cBrain, the Danish company behind the ministry's F2 case and document management system, was tasked with integrating Collabora into it.

The ministry presented it as a test from the beginning and said that a wider rollout in the department depended on the results. Days later, it told Version2 (in Danish) that no decision to phase out Microsoft had been made and that the plan was not to replace Windows with Linux.

Ingeniøren reported in December (also in Danish) that testers had complained about documents from other government agencies that opened blank, track changes that were difficult to accept, and line spacing that appeared odd.

Its story was based on ministry and cBrain correspondence obtained through a records request, and it covers the months up to an evaluation workshop in October, where testers were to say whether they wanted to keep using Collabora.

Later that day, a second story had Frejdie Søndergaard-Gudmandsen, cBrain's head of product management, calling the bugs normal for a pilot. It was also reported that the ministry had decided to continue into 2026.

Two versions of why it stopped

The ministry's press officer told Ingeniøren in October that the election call interrupted the project, that it was not evaluated as planned, and that staff don't use Collabora today and have no plans to.

Collabora's statement gives a different cause. It says the June reorganization ended the digitalization ministry as a separate ministry and left the pilot without a home before the evaluation was done, and that this shouldn't be characterized as a finding against the suite.

The election was called in February, and the new government that took office in June merged the digitalization ministry into the research and education ministry.

Collabora adds that the first phase worked, that its engineers resolved the tickets testers raised, and that it paid for that work itself. It also says the experience is carrying into work with other Danish authorities and cBrain, and that formal assessment was never finalized.

On October 27, the research, education, and digitalization minister Christina Egelund said on DR's P1 Morning show that she knows little about why the project ended without a proper evaluation and pointed out that the decision was made under the previous minister.

Where our earlier coverage ran ahead

Our June 2025 article said, citing Politiken's report, that the ministry was moving from Office 365 to LibreOffice, with full adoption expected by autumn. The ministry later said no decision to phase out Microsoft had been made and described a smaller volunteer pilot.

Politiken soon corrected its own article, and we fixed the Linux claim at the time, but our headline and the autumn timeline stayed as they were.

Amutable has opened signups for design partners to test an immutable, image-based system that ships every change as a signed update. Built by the engineers behind systemd, it uses hardware attestations to verify each machine’s state before allowing deployments.

Gregory Kurtzer's Ginux is a brand-new Rust kernel that runs Linux binaries unchanged despite not being Linux underneath. An AI agent generated 139,000 lines across 11 days, proving a Linux-compatible kernel can emerge from scratch without needing to fork the original.

BeamFS landed on the kernel mailing list this week as an RFC with warnings about existing data loss bugs. Still untested in real radiation environments, it prioritizes detecting corruption over silently serving bad data.

FreeBSD founder Jordan Hubbard built Daemonarchy in under two hours using GPT Astra 6 as an answer to the many requests he received that said FreeBSD deserved an installer as slick as Omarchy.

Tesco is ditching VMware for Proxmox in its stores after Broadcom stopped supporting perpetual licenses and demanded 175% higher prices. A newly posted job suggests a gradual migration off VMware by the end of 2027.

This edition of FOSS Weekly is supported by Surfshark, a privacy company based in the Netherlands.

🧠 What We’re Thinking About

Mirko Brombin from Bottles says he is pushing Microsoft Office onto Linux because real-world Linux users often need it, regardless of LibreOffice’s merits.

🎃 Tales From the Linux Crypt

Halloween is here. While I will revisit Agatha Christie's Hallowe'en Party, I will share some morbid, scary but fun articles for you to read 👻

Linux is scary place, isn't it? There are zombie (processes), kill (command) and daemons.

Jokes apart, there are several dangerous Linux commands like rm -rf / or the fork bomb that you should never run on your system.

Since Halloween is the festival of remembering the dead, let's take a look back at some of the loved Linux distros of the past that no longer exist.

Since I started my Linux journey with Ubuntu 17 years ago, I have seen several ambitious projects killed by Ubuntu. Can't really blame them, but I would have loved to see the convergence dream being a success.

Although the discontinuation of scp command is greatly exaggerated, there are quite a few Linux commands that have been deprecated over time.

Resurrecting the dead is sort of a tradition in open-source world. Linux world is moving towards Wayland and so there are a few projects that are working to keep Xorg alive. And then here are some tools that got new life in the form of "new generation" version. If they were resurrected today, perhaps that can be renamed to "gen z" versions ;)

Those are not the only ones, there were several prominent projects that changed their license, and hence a new open source project rose to take their place. Open source software never really dies.

Halloween Makeover

Above is the Kitty terminal with a special Halloween makeover. You can get the config file here.

If you would like to give your Linux system a little bit of Halloween touch, we have a few references for you. For Hyprland users, there are a few dot files here.

And a demo video showing how GNOME desktop can be customized with a "scary look".

👷 Homelab and Hardware Corner

I tested a tiny, inkless, portable printer on Ubuntu 26.04. Interesting concept for a printer.

In other news, ArmSoM’s Sige7s squeezes the RK3588S chip into a Pi-sized board with a 6 TOPS NPU, Wi-Fi 6, HDMI 2.1, and up to 32GB RAM.

A modular USB-C hub combining Framework-compatible Expansion Cards with tools for measuring, powering and prototyping electronics is being crowdfunded.

✨ Apps and Projects Highlights

massCode is an open source note-taking app but specifically programmers for saving code snippets. The app has revamped recently with new features, including the MCP option.

📽️ Videos for You

I installed the openKylin distro from China and tested its "air gesture" and integrated AI agent features. You can watch it in action in this video.

💡 Quick Handy Tip

You can explore the contents of compressed archives (like ZIP, TAR, or GZ) directly in Dolphin as if they were regular folders:

1. Click the three horizontal lines button, go to Configure, and select "Configure Dolphin..."
2. In the left sidebar, click "View" and stay in the "General" tab.
3. Here, check the box for "Browse compressed files as folders", then click "Apply" and "OK."

🎋 Fun in the FOSSverse

Okay, just a couple more Halloween stuff, I promise 😄 Here's some fun Halloween-themed Linux puzzles and riddles for you to enjoy.

📆 Ubuntu 6.06 was the only release that got delayed and was not released in April (XX.04) or October (XX.10). Otherwise, Ubuntu always follows strict deadlines.

🧑‍🤝‍🧑 From the Community: The recent PHP vulnerability scare affecting programs globally has got our resident FOSSers talking.

openEuler has two fresh releases out. 24.03 LTS SP4 arrived in July 2026, and the short-term 26.09 innovation release followed on October 1.

We covered openEuler 24.03 LTS when it came out in 2024. Two years later, that LTS version is still getting new features, and SP4 is the latest round.

The releases have plenty of features of all kinds. Some target SuperPoD infrastructure and specialized AI accelerators, you know, enterprise features. So this article focuses on the features you and I can explore on a regular PC or virtual machine.

I am going to check on the AI tools, the desktop bits and the under-the-hood changes. I installed them in GNOME Boxes to see what actually works on a regular machine.

LTS, service packs and innovation releases

openEuler has two release tracks: long-term support (LTS) and innovation releases. The community is continuing to develop the 24.03 LTS series through service packs, with the next major LTS release currently expected in 2028.

Instead of new version numbers, an LTS gets feature updates through service packs (SP1, SP2 and so on). Unlike Windows service packs (if you remember it), these are proper feature releases, not just a bundle of patches.

Similar to Ubuntu's six monthly releases, these "Innovation releases" introduce newer technologies on a shorter support cycle. openEuler 26.09 is scheduled to receive support until March 2027.

What's new in openEuler 24.03 LTS SP4

SP4 stays on Linux kernel 6.6, but openEuler keeps backporting newer upstream features into it.

AI-assisted diagnosis and tuning

The AI addition most relevant to regular users is intelligent diagnosis and tuning. An Intelligent Diagnosis Agent works with a Known Issue Analysis Agent to find a problem, work out the root cause and suggest a fix.

The tuning side collects system metrics, finds performance bottlenecks and gives you optimization suggestions. In theory, that saves you hours of staring at top and perf output.

The AI features are delivered through openEuler's Witty framework, which can be installed from the distribution repositories using the package witty. Once installed, you launch the interface with witty or opencode and interact with the diagnosis and tuning agents through natural language prompts.

Unlike the first-generation EulerCopilot experience, which was closely tied to openEuler's own cloud AI services, Witty supports both local models and a range of OpenAI-compatible APIs, making deployment significantly more flexible for users outside China.

There's also a lightweight Agent sandbox for running code and tool calls from AI agents in isolation. Considering how many people now give AI agents shell access, that's a sensible addition.

Kernel changes worth knowing

Under the hood changes that you won't see.

The biggest addition is sched_ext, the eBPF-based scheduler framework that landed upstream in Linux 6.12. You can load custom CPU schedulers at runtime without recompiling the kernel. If one misbehaves, the kernel quietly falls back to the default scheduler.

FUSE file systems can now talk to the kernel over io_uring. That should help performance for FUSE-based tools like sshfs.

There's also a programmable page cache that lets applications set their own file caching policy. It's meant to speed up loading large AI models, which is usually limited by disk I/O.

ARM users can now choose a 64K page-size kernel at install time. The 4K kernel stays the default, so nothing changes unless you opt in.

The enterprise stuff

The rest of SP4 is aimed at enterprises and data centers. There's NPU compute slicing, faster recovery for AI inference containers, networking and reliability features for SuperPoD, and confidential computing improvements for Arm CCA, TaiShan and Hygon systems. openEuler's own compiler builds (LLVM, Go and BiSheng JDK) also got updates.

What's new in openEuler 26.09

26.09 ships with two kernels: Linux 6.6 for servers and Linux 6.18 for DevStation, openEuler's developer desktop edition.

DevStation is where most of the desktop-relevant changes are and that's what I tried in the VM.

A new installer and a user-space package manager

DevStation uses a new installer called "heolleo". I almost read that as "Hello Leo" 😄

It uses a SquashFS-based image, can install from a local ISO and works fully offline.

The more interesting addition is epkg, a user-space package manager. It installs software in your home directory without touching system packages, and you can keep separate environments for different projects. It supports multiple package sources too, which makes it sound a bit like Conda or Nix.

PolyMind: a self-hosted AI agent platform

DevStation also bundles PolyMind, a self-hosted platform for AI chat, agent workflows and managing multiple models. It connects to MCP services and AI agents through its agentd service.

SkillHub and keeping agents in check

openEuler also launched SkillHub, a place to find, install and share AI agent skills. Each skill gets an automated security assessment. openEuler is upfront that this helps you judge risk but doesn't guarantee a skill is safe, which is a fair warning.

AgentCensor is the one I find most interesting and useful going by its description. It sits between an AI agent and your system, logs tool calls, file changes and system calls, and can allow, block or ask for your approval before the agent runs something.

SCCS tackles token bloat in long-running agents. It replaces large tool outputs in the context window with short references and fetches the full content only when needed. The rest (ram-a-kv, MemFabric and ModelFS) are for optimizing inference servers.

ModelFS is a model startup acceleration feature that uses a programmable page-cache framework underneath, allowing caching behaviour to be tuned for AI workloads while remaining compatible with existing applications.

Getting openEuler

You can get both from the official download page. There are several choices, from servers to embed devices. For normal usage, go for the DevStation. That's the desktop version and that's the one that comes with all the above mentioned AI tools.

Things have certainly improved in the past two years since we last covered openEuler first. The focus is on providing AI and that's not surprising. We are in 2026 and AI is everywhere. An operating system focused on developers has to provide AI at its core.

Have you tried openEuler ever? What's your experience with it like? Share in the comments please.