Welcome to Local AI Weekly #4.

"Local AI" keeps meaning different things these days. Sometimes the model runs on your computer. Sometimes only the app does, while the model and your chat still run elsewhere.

There are lots of such "local AI" tools that are not really local. Before you pick a local AI tool, ask three things: where does inference happen, what account or network service is still required, and what does the license let you do? The best local AI tool is the one that doesn't need any of that.

🧪 On my bench

Team It's FOSS has moved from Discord to Buzz for internal chat. Buzz, from Jack Dorsey of Twitter fame, is a decentralized communication tool for humans and agents. Agents can run on remote servers or a local harness over ACP.

It has quirks. Clipboard screenshots won't paste into chat, and desktop notifications only fire for direct messages. Manageable for now.

🔍 Discover AI tools

Kubutu developer Rick Timmis is working on Klara, a local desktop AI assistant for KDE Plasma. The idea is to let you control the desktop via voice input. It is a work in progress for now.

OpenMuse is an MIT-licensed personal-agent app with a browser worker, durable tasks, and an optional Docker-based Linux computer. You can host it yourself, but setup still needs a CopilotKit Intelligence key, and open-ended tasks default to cloud providers. You can point it at an OpenAI-compatible endpoint, but there's no documented native Ollama path. Self-hosted software, not an assured offline agent.

🎫 Get MCP Certified

The Linux Foundation now offers AI certifications. The Model Context Protocol Associate one could interest you if you're building MCP integrations, or want an AI credential on your resume. I plan to take it just for the sake of learning new skills.

📡 Open Model News

OpenDecider is a more modest use for local models: answer bounded questions, like which queue a ticket should go to, instead of writing long replies. Its nano model is about 400M parameters; the 4B small model is a Qwen-based adapter. The author says both were distilled from larger teachers, and reports 2.0 GiB for nano and 8.9 GiB for small in the tested setups.

Basically, you don't always need a giant model to route routine requests. A small student model can be the triage step ahead of a slower agent.

👀 Big Tech Watch

NVIDIA's September PAIR announcement also promises easier local-model setup in Hermes and OpenClaw. The one-click Hermes path launched on Windows, with Linux "coming soon." Don't confuse that future path with the Linux PAIR beta available now.

NVIDIA also quotes up to 1.9x throughput from llama.cpp optimizations on an RTX 5090. That's a vendor number on specific hardware, not a speedup I'd expect on yours.

🗂 AI Jargon: Distillation

Imagine you have a giant, super-smart teacher who knows everything about the world. This teacher has a massive brain, but its so big that it can only stay inside a giant school building.

AI distillation is like that big teacher sharing all their secrets with a little kid (the student).

Instead of making the kid read millions of textbooks, the big teacher says: "Don't worry, just watch how I solve these puzzles, and listen to how I think."

The little kid watches closely and learns the teacher's smart shortcuts. Soon, the kid becomes almost as smart as the teacher, but with a much smaller brain!

You can learn more about distillation here. And you will see that teacher-student is kind of official term in this context.

😂 Meme

When the open-weights drop looks a little too familiar...

⚡ Quick Tip: Back up your Hermes agent before you need to rebuild the harness

Last issue: ollama ps, to see if your model was really on the GPU. This week, make sure you could rebuild the setup around it too.

If you run Hermes, run hermes backup. It writes a ZIP of your Hermes home, config and state included, restored later with hermes import path/to/backup.zip. hermes backup --keep N caps retained backups, and a script-only cron job runs it on schedule without starting an agent.

Then move one encrypted copy off the machine. That archive can hold credentials, sessions, memory, and config, so don't drop the raw ZIP in Git, even a private repo. It won't be wise.

See you next week.

Announced via a two-part series of blogs, Microsoft has moved WSL containers (WSLC) out of public preview and into general availability. Let's take a look at what it offers.

The new release comes with wslc.exe, a dedicated command-line tool for Linux container workflows on Windows. It ships with a built-in alias, container.exe, for those who prefer that syntax.

A Windows API also ships alongside it, giving native Windows applications a way to spin up and manage containers directly from code, as well as some new commands that include wslc events for tracking container activity, and --mount and --stop-timeout flags on create and run operations.

Networking is handled through a new model called Consommé, where container traffic leaves the virtual machine as Ethernet frames and is picked up by a Windows process running under the calling user's account. That process handles DNS, routing, and port mapping, letting traffic pass through VPNs and firewalls like any other Windows process.

For organizations, Microsoft Intune has gained two new settings specific to WSL containers. The first lets administrators enable or disable access to the WSLC feature entirely across managed devices. The second is a container registry allow list, which restricts image pulls to a defined set of approved sources.

Microsoft Defender for Endpoint's WSL plugin has also been extended to cover container activity. It can retrieve process, file, and network events from inside WSLC, connecting them back to the Windows host.

What is WSLC?

First you have to know about WSL, which stands for Windows Subsystem for Linux, that lets developers run Linux environments directly on Windows without needing to partition their drive or setting up a separate Linux machine.

Since WSL 2, it has shipped with a real Linux kernel inside a managed virtual machine, giving users access to Linux tools, distributions, and command-line workflows from within Windows.

WSLC extends this further by adding a dedicated layer for creating and managing Linux containers within that same WSL environment, making containerized workflows a native part of the setup.

It also separates container operations from the main WSL service by routing them through a dedicated child process, wslcsession.exe, which runs under the current user's account. This keeps each session isolated and container operations in a less privileged state than the WSL service itself.

For developers already working inside WSL, this means containerized applications can run in the same environment without reaching for a separate tool. The Windows API exposure also means native Windows applications can interact with containers programmatically.

Microsoft's WSLC architecture deep dive is a must-read if you want to know more.

Get it now

Running wsl --update in your terminal pulls in the latest WSL release, which includes WSL containers. Once updated, wslc is ready to use.

You can use these new commands to manage your containers. 👇

For the full changelog and access to the source, head to WSL 3.0.1's release page on GitHub.

When we heard about the Netherlands' DAWO initiative, a formal mandate to replace outside software across Dutch public institutions with a NixOS-powered platform, we asked ourselves: How could France miss out on this?

Turns out, they didn't. DINUM, the country's Interministerial Directorate for Digital Affairs, has been running its own NixOS-based workstation setup on internal machines for months now.

Sécurix and its NixOS Bet

DINUM has developed Sécurix, a security-hardened NixOS setup aimed at system administrators. It builds on top of NixOS without forking it, applying the security guidelines published by ANSSI, France's national cybersecurity authority.

The project is MIT licensed, lives under the cloud-gouv GitHub organization, and is currently in alpha, with v0.20.1 being its newest release having been introduced just a few hours ago.

According to Emilien Ercolani of The Stack, who attended a DINUM briefing in April, they are targeting implementation across ~250 of their staff's workstations. This comes after a successful pilot run across 70 machines.

In its current avatar, Sécurix uses FIDO2 hardware security keys as the main login method, supports TPM2 and YubiKey, and leans on NixOS's reproducible build model.

Keep in mind that this is not meant to be a Linux distro but rather a foundation from which workstations can be operated and managed centrally.

Alongside it, DINUM has published Bureautix, a companion reference template for regular office workstations. It ships with KDE Plasma, LibreOffice, ONLYOFFICE, and WPS Office.

The docs describe it as a "dummy example" that organizations are meant to fork and adapt privately, not a separately running system.

Why is NixOS the go-to?

The project itself has a European origin. Eelco Dolstra created it at Utrecht University in the Netherlands, and the NixOS Foundation is a Dutch nonprofit. For European governments, the appeal is obvious.

Mainstream Linux distro offerings like Fedora trace back to Red Hat, which is owned by the US-based IBM; openSUSE is tied to the Luxembourg-based SUSE; and Ubuntu is from England's Canonical.

The underlying approach here seems to be avoiding projects that are based in foreign countries that could be forced to comply with the whims of their leaders.

From what I can see, France, the Netherlands, and Denmark are each taking a different path to the same place. In the end, if one wants control over their infrastructure, vendor-neutral and reproducible is the way to go.

Suggested Read 📖: postmarketOS is no more; not in that sense! It has undergone a rebrand.

openSUSE is known for shipping four main Linux distributions. Leap is the stable, fixed-release option built on the same source as SUSE Linux Enterprise (SLES). Tumbleweed is the rolling release for those who want the latest packages. MicroOS is the immutable, rolling variant that tracks Tumbleweed.

And finally there's Leap Micro, the immutable, fixed-release counterpart to Leap, built for container hosts, edge devices, and virtual machine workloads. Its job was to bring immutability as well as stability to openSUSE without affecting Leap itself.

That separation is ending as the project has now decided to ship immutable Linux as a separate mode inside the main Leap installer.

A new mode

Leap Micro existed as its own product because immutable systems required a separate installation path. With Leap 16.1, the Agama installer has gained an "Immutable" mode option that can be toggled on during setup.

If you go with that, your system will be configured to have a read-only root filesystem with transactional updates and snapshot-based rollback. The technical foundation is identical to what Leap Micro offered; the difference is that it now lives inside Leap rather than alongside it.

System updates run through transactional-update, which applies changes in a new snapshot before the reboot, letting you roll back when a bad update causes issues. For installing new software, Podman and Distrobox cover containers, while Flatpak handles desktop apps.

Don't worry, you can still opt for the usual desktop experience if you choose the "Standard" mode.

And thanks to this, there will be no Leap Micro 6.3 or 7.0 releases as they have decided that Leap Immutable is the successor for anyone running container workloads, edge deployments, or setups that benefit from atomic updates.

For existing Leap Micro 6.2 users, the project offers the experimental openSUSE Migration Tool as a path forward.

Anything else?

As this is a new Leap release, there's a lot packed in other than the new mode. You get a range of upgrades that include KDE Plasma 6.6 with its OCR support in Spectacle and accessibility upgrades, as well as GNOME 48 with its HDR support and the new default audio player.

The latter is the same version that shipped with Leap 16.0 but with a gnome-shell bug fix patch bundled in. The openSUSE devs are staging the GNOME upgrade gradually over the next couple of releases rather than jumping versions all at once.

You can expect the stable release of Leap 16.1 to land the same day as SLES 16.1.

Get the RC

For now, only the release candidate (RC) is available, with the project committed to publishing one RC build a week before they reach Gold Master Candidate (GMC) status.

You will find RC images for x86_64, aarch64, ppc64le, and s390x from the Leap website.

Firefox has a new look that brings it up to 2026 web browser standards, where having a clean, minimal user experience is a growing phenomenon.

Called "Nova," the redesign leaked before an official announcement was ever made, with Mozilla playing catch-up a few months later via a technical blog outlining how they were approaching this major transition.

That transition turned out to be more work than the mockups made it look.

Six months after the leak

When the internal design files were made public in March, the mockups looked unlike any Firefox build before it. The tab bar, toolbar, and address bar were pulled into a single floating island at the top, with a visible gap between them and the page below.

Web content got the same treatment in those designs, sitting inside a rounded container lifted away from the browser's edges rather than running along with them. But the thing is, neither idea made it into the final release.

What we see in Firefox 157 is more refined. Tabs, menus, and the address bar are noticeably rounder, with a subtle gradient on the active tab that adds warmth to the interface.

Come, have a look

I pitched Firefox 156 against 157 to see what changed, and I came out impressed.

Opening a new browser tab showed me where the most effort was put, with the tab bar being visibly rounder and the address bar corners being softer, giving the interface a very polished look.

Before this, I was already beginning to dislike the flat/bland look Firefox had; it bugged me whenever I would return from using another browser like Vivaldi or Chrome.

You see, both of them have already embraced a pill-shaped design philosophy that I quite like.

Loading up a website was more of the same, with the slimmer footprint of the interface leaving the webpage more real estate to use, and the toolbar has fewer dividing lines between its elements than before.

The settings menu has also been refreshed. In 156, the sidebar was presented as a compact list with fairly simple content panels. In 157, those panels sit inside rounded cards, with sections that give each settings area more visual room and better readability.

What else's new?

Tab Groups has changed considerably since I last used it. The Nova redesign has given it a more complete feel compared to when the feature first landed in Firefox 138.

Back then, groups showed as colored labeled strips in the tab bar, and management options were limited to naming and collapsing. While the feature worked, it felt like an early implementation.

Now, tab groups sit more naturally in the tab bar, and switching between them or collapsing one out of the way feels less clunky and more intuitive to use.

The "List all tabs" button has been changed too, featuring a different icon (instead of the downward arrow) and visual style that ties well into the overall pill-shaped design. You can use it to get a full view of every tab across all groups at once.

A feature that was stripped out back in 2021 has made a return. Compact mode is now available in Settings > Appearance under the "Window density" selector in addition to the Automatic, Standard, and Touch options.

You will also find many new themes that can be installed via the Add-ons manager, as well as a new Classic/Smart Window switcher button on top-right for quickly switching between traditional and AI-powered browsing.

You can disable it

If you are not impressed by what Nova has to offer, then you could disable it to get back the old Firefox experience, and it's quite simple to do. In the address bar, type about:config, and accept the scary-looking warning.

Now, go into the search bar and type this "browser.nova.enabled" without the quotes. The final step is to either double-click on the entry or on the "Toggle" button to move Nova's activation state to "False."

That's it; you have successfully reverted to the earlier design.

Download Firefox

This is the obligatory installation section of the article, where you will find that Firefox can be downloaded from the official website. Though, when I was testing the redesign, I had to rely on Mozilla's FTP mirror to get the Firefox 157 release early.

I went with the package meant for my test setup: linux-x86_64 > en-US > firefox-157.0.deb

If you are on Debian or any of its derivatives, like Ubuntu, you can add the APT repository to get it installed.

And for existing users, by the time you read this, the various mainstream Linux distributions (e.g., Ubuntu/Fedora) should already be serving this redesign as part of a routine system update.

Suggested Read 📖: Smart Window users now get to use French AI.

postmarketOS, the Linux-based operating system that many of you turn to for reviving your old smartphones, has undergone a name change, and it is now called "Nura."

This was announced during the project's conference recently, where the rebranding closed a long-running process that started last year in March, drawing more than 300 name submissions from the community.

It took a dedicated selection team and a trademark review aimed at accommodating languages from every continent to ensure that the new name didn't have any offensive meanings across different cultures.

postmarketOS had to go

The old name worked against the project in ways that compounded over time.

It was long, difficult to pronounce across many languages, and its mid-word capitalization had no obvious logic for anyone coming to it fresh. The project says that people were mischaracterizing it as "PostmarketOS," "Postmarketos," and "post-market OS."

Then there was the problem of scope; initially, when the project was focused on reviving old smartphones, it was an appropriate moniker to flaunt. But when PINE64 began shipping phones with postmarketOS pre-installed, it became a liability as they couldn't trademark a descriptive name.

That left the project with no legal standing to protect its identity against misrepresentation.

With Nura, a trademark application has already been filed, though it has not yet appeared in public trademark databases so far (I searched USPTO and EUIPO).

What's Nura?

Nura is short for Nuraghe, which are granite constructions scattered across Sardinia, Italy, with some of these being traced as far back as 3000 BC. Many of those structures are still standing, giving us a glimpse of what the Nuragic civilization built back in the day.

The project draws inspiration from that, seeing longevity as a direct match for its own goal of keeping smartphones useful for a decade or more.

The name itself came from Davide Depau, who submitted it during the project's open call for community contributions.

The selection team, consisting of Pablo C. Gómez, Pan Ortiz, Ranny Bergamotte, and Sriram Ramkrishna, worked through more than 300 submissions and narrowed the field to four finalists.

The team voted, and Nura won.

They also went for the .eco domain knowingly, as it is reserved for organizations committed to positive environmental change and comes with enforceable policies against greenwashing.

The project's current status

The new website presents a very straight-forward look at Nura's status, mentioning that it works best for people willing to learn how it works, contribute to testing, or help with development.

Its current release is v26.06, and for anyone getting started with Linux on mobile, the project points to second-hand Snapdragon 845 or Snapdragon 410/412-powered devices as the most sensible entry points.

Currently supported devices include the Fairphone 4, OnePlus 6 and 6T, PINE64 PinePhone and PinePhone Pro, Purism Librem 5, and several older Samsung and Xiaomi devices.

Suggested Read 📖: Your search for a Linux phone ends here.

The man leading GhostBSD, Eric Turgeon, has come up with Mocka, a new desktop for this BSD distribution that would eventually replace MATE. So far, he has unveiled the dock that will go into it, as well as some preliminary concepts and specifications that show us where Mocka will head next.

Existing users of GhostBSD don't need to do anything now since nothing has changed on their machines. All of this is still under heavy development after all. 🤷

Grammar nazis won't like this

In his search for projects named "Mocha," the caffeinated drink, Eric made a typo and ended up naming his new creation "Mocka," where every component has been built from scratch, taking in nothing from the MATE project.

He plans to replace one component at a time, reverse-engineering what MATE already provides until full feature parity has been achieved. Once this is complete, I assume that GhostBSD will default to using Mocka as its default desktop environment.

And this is not his first time carrying out such a major change. In February, he came to the conclusion that MATE was not ready for Wayland and GhostBSD would switch to XLibre to keep its X11 session intact.

That line of thinking also checks out with what Eric's doing now.

The desktop's first component, a taskbar-style dock applet for the MATE panel called Mocka Dock, is already out now as an alpha for early testers. It lets you open apps, view their thumbnails, bring them forward into focus, and manage multiple windows of an app—basically doing what a dock worth its salt should do.

People who have the skills can build it from source right now to see what it offers, while the rest of us will have to wait for it to show up as an GhostBSD package update. Though it is not becoming the default experience yet, you will have to manually set it using the "Add to Panel" button on the taskbar.

What's planned?

If you take a look at Mocka's GitHub page, you will get a better understanding of what's being worked on apart from the dock. There's Mocka Menu, which has a proper PLAN.md file that lays out what work has already been completed and what's to be done.

The document shows us how a meson build and schema work has been done, with a long list of work yet to be carried out. This would give Mocka the ability to show a menu with listed items, the ability to handle search and keyboard inputs, integrate with the dock, and much more.

Eric has plans for a settings app too. Once complete, it would give users a single-access point for managing their GhostBSD machine's desktop appearance, network config, power management, keyboard layouts, and other UI-focused options.

There's also a separate repo that hosts the code for Mocka's official website. It is considerably behind in terms of development than the menu component we saw earlier and only contained placeholder content at the time of writing.

The project is hoping to ensure that these new Mocka components remain compatible with MATE so that users can mix-match during this transitional phase to a full Mocka desktop on GhostBSD.

For the first time, there's a way to carry out software certification on AlmaLinux, and the hardware suite has been rewritten alongside it, with test results from either certification path landing in the same public catalog.

Don't worry about the cost to validate or the uptime of the machine itself. Certifications are free, and the hardware checks no longer require you to install the distro on your setup.

Validating, how?

In this particular situation, software certification is basically a listing plus the confirmations that follow it. A publisher adds the product and picks which major versions of AlmaLinux it runs on.

The confirmations come from the people running software on their own hardware across various AlmaLinux releases, though they will need to sign up for an account if they want to post their findings.

Hardware certification is the other half that got some attention.

The alma-certify tool has been rewritten, with the developers claiming it can now achieve sub-10 minute certification runs on most hardware.

Another improvement is that alma-certify doesn't require an AlmaLinux installation anymore because it can work off live media or from an installation already on the machine. There's also a new machine registration flow that makes use of QR codes.

If you didn't know, this tool records what's in the machine, all the way down to drivers and firmware, then starts up a bunch of short checks that decide whether it passes.

Those checks cover a lot of ground, starting from computation and memory errors, storage health, networking drivers, to kernel state, virtualization, GPU, and peripherals.

Tracking results

Everything ends up in the catalog, and each entry is a result anyone can analyze. It holds the systems, components, and software that have been validated, alongside the benchmark figures from the published test runs.

Every major AlmaLinux release is tracked on its own, so provenance can be proved, and everything in the catalog is easily searchable. You can even make use of the free, read-only API to collect data in bulk.

Jonathan Wright, the Infrastructure Lead for AlmaLinux, directed his attention toward potential testers, saying that:

Two things would make all of this worth it. A lot of people and organizations want proof that AlmaLinux runs on the hardware they already own before they’ll give it a try, and now they can get that proof themselves.

Every result that gets submitted makes the case to hardware and software vendors that supporting AlmaLinux officially is a low-effort thing to do.

So if you have hardware sitting in front of you, or software you rely on every day, go tell us that it works. Validations stack, so adding yours to something already listed helps just as much as being the first.

In the end, this will only work if vendors and individual users take the effort to run tests and manually post them on the portal.

Suggested Read 📖: Ubuntu is tightening its kernel SRU cycle to two weeks.

When the U.S. government imposed sanctions on the International Criminal Court (ICC) in early 2025, the consequences were far-reaching yet quickly enforced.

The restrictions cut off the ICC chief prosecutor's access to Microsoft services, including email, followed by a freezing of his personal and institutional bank accounts in affected jurisdictions.

While this was happening, the Dutch government was watching, coming to the bitter realization that access to essential services depended on decisions made in the U.S.

Hence, their response to this is DAWO, short for Digitaal Autonome Werkomgeving Overheid, or Digital Autonomous Work Environment for Government in English.

It is a framework that looks to tackle the government's digital stack, covering the OS, an office suite, collaboration apps, cloud services, IT management tools, and AI, with all of it being built around a NixOS foundation.

How it came to be

The DAWO initiative became an official government mandate in July, when the Interdepartmental Commission for Government Business Operations (ICBR) issued a directive for a standardized, sovereign digital work environment.

The three major IT service providers of the Dutch central government, SSC-ICT, DICTU, and DUO-ICT, are tasked with executing this major undertaking.

Development is handed by a team that is associated with the Ministry of the Interior and Kingdom Relations (BZK), with some notable people like Victor Gevers, Rutger Putter, and Bram Buijs being named.

Why go with NixOS?

Nix, the package manager NixOS is built around, handles configuration through its own functional programming language. Each package installs into an isolated directory with immutable, signed contents.

That makes a working setup reproducible across machines. For a rollout spanning dozens of agencies, that reproducibility adds up, resulting in substantially less repeated configuration work that would cost time and money.

Rutger estimates around 80 to 90 percent of a configuration carries over between deployments. NixOS also runs on hardware Windows 11 cannot, which is why the current pilots (more details below) are being run on decommissioned laptops that would otherwise be out of service.

For handling the code, this initiative is following a similar approach to the country's earlier Forgejo migration, where they moved their code hosting away from GitHub and GitLab.

All of DAWO's code is hosted on a Forgejo instance, with a separate Codeberg instance acting as the entry point for contributors. And, to answer the question asked earlier, NixOS was chosen because it is a Dutch project, with no company owning or controlling its development.

Before settling on it, the DAWO folks were considering openSUSE and Fedora. Both were disqualified due to their commercial ties, openSUSE to SUSE and Fedora to Red Hat, which in turn is owned by the US-based IBM.

Suggested Read 📖: Garuda Linux has begun pushing towards NixOS.

Things are looking up

Eight municipalities are running trial programs via the Association of Dutch Municipalities (VNG).

The ICC incident is what makes this different from other government-level open source pushes, as Dutch officials were given a first-hand reminder of what vendor dependency could cost them.

Of course they are not alone in doing so. Germany's Schleswig-Holstein went through something similar after leaving Microsoft. Part of what it used to spend on licenses now goes towards supporting open source projects.

Denmark and France have made similar moves of their own, where one is ditching MS services for their road traffic authority and the other is pushing a Linux switch across ministries, with 80,000 health insurance employees already moving to domestic tools.

As for where DAWO goes from here, they intend to release the first stable version sometime next year, taking in feedback from all the concerned parties (e.g., the municipalities).

Via: Tweakers

You get an SBC and start building projects on that. You start with a vision project. Then the next project idea needs audio. Perhaps the next one requires a bunch of sensors. Usually that means a stack of HATs on the GPIO header or getting a new SBC.

Vicharak is trying a different route with the Axon-Lite. It's a tiny ARM SBC whose interface modules are swappable, so you change the module to match the job instead of changing the board.

This "reconfigurable computing" pitch is what makes this ARM board slightly different from the rest.

Vicharak already offers a FPGA-based Vaaman and an RAM-based Axion line. The Axon-Lite is the smallest member of its Axon line. So they may be relatively new in this field, but they are not total novices.

Axion Lite Specifications

Let's see what you get on this SBC:

• SoC: Rockchip RK3576 (4× Cortex-A72 + 4× Cortex-A53)
• AI/graphics: 6 TOPS NPU, Mali-G52 MC3 GPU, 16MP ISP
• Memory & storage: LPDDR5 (2GB-8GB), eMMC 5.1 (32GB-128GB), UFS 3.1, SD card
• Modularity: swappable interface modules + board-to-board expansion
• Connectivity: Wi-Fi 6, Bluetooth 5.2, Gigabit Ethernet
• GPIO: Raspberry Pi-compatible 40-pin header

As you can already see, the highlight here is the modularity. The board uses what Vicharak calls its modular connector interface, and the idea is that you snap on a different interface module as needed.

The company shows four so far: Voice Processing, Smart Sensing, AI Data Hub and AI Vision. On top of that, there's board-to-board expansion, so you can stack a daughter board rather than swap.

There's a but worth keeping in mind. A modular connector is only as useful as the module lineup that actually ships for it. And if this is a proprietary interface, not a standard one, so you will be dependent on whatever expansion board Vicharak is offering.

Under the modules sits a Rockchip RK3576. It's an octa-core chip that slots in above the entry-level RK3566 and below the flagship RK3588, which makes it a sensible middle ground for edge work.

Vicharak lists a 6 TOPS NPU alongside it, the same NPU figure you get on much pricier RK3588 boards, which is plenty for running vision and other on-device AI models locally. The company also claims the SoC is built on a TSMC 5nm process.

Memory is LPDDR5, and storage is flexible with eMMC 5.1, UFS 3.1 and a microSD slot all supported. The inclusion of faster UFS on a board this size is a nice touch.

For vision builds, the I/O is generous. You get a 4-lane CSI plus additional 2-lane CSI paths for cameras, a 4-lane DSI for displays, a 16MP ISP, micro HDMI 2.1 and USB-C with display output. Networking covers Wi-Fi 6, Bluetooth 5.2 and Gigabit Ethernet, and there's a Raspberry Pi-compatible 40-pin GPIO header so existing add-ons and pinouts carry over.

Vicharak also mentions on-board battery charging circuitry, which hints at portable, untethered builds.

Software and userbase

Vicharak is pitching the Axon-Lite as a Linux-first board, with a familiar Debian-style environment and the usual apt-driven workflow. Nothing exotic to learn here, in other words.

On the tooling side, the company lists the tools you'd expect for this kind of board: VS Code, Git, Python, Docker and Node.js for general development, and ROS 2, OpenCV, GStreamer, FFmpeg, Jupyter, TensorFlow Lite and ONNX Runtime for the AI, robotics and vision crowd.

Between that and the swappable modules, the target audience is quite spread as it is useful students, hobbyists and product teams prototyping edge AI and vision projects.

Pricing and Availability

Vicharak is pitching the Axon-Lite as a Linux-first board, with a familiar Debian-style environment and the usual apt-driven workflow. Nothing exotic to learn here, in other words.

From students tinkering with AI and embedded systems to hobbyists to product teams prototyping edge AI and vision projects, the target userbase for the board is quite spread.

Axon Lite is available for pre-order for the 4 GB and 8 GB RAM variants with a price tag of Indian rupees 15,930 (around US $165) and 18,299 (around US $200). The board should ship in January 2027. Please check for shipping and custom duty charges.

It is always interesting to see alternative single board computers, specially if they have something different to offer.

The AI doesn't necessarily run on supercomputers (yet), it runs on clusters of hardware pieced together. This is where SuperPoD comes into picture. A SuperPoD is a tightly coupled scale-up system that connects many compute nodes so software can use resources across them more like one system.

So far, the operating system that ties them in a single machine has largely been proprietary. openEuler wants this to be open source, and its latest release shows the push for that.

Need for cluster OS

Linux provides the base for an excellent operating system. But mostly, the OS is confined to one machine. It schedules the memory, the CPU, and the devices inside a single box, and if you want more, you add another box and wire them together over the network.

This has worked so far, but the AI ecosystem demands more than that. Models now run across racks of machines that have to behave like one giant computer (SuperPOD, again), and the network in between becomes the bottleneck for performance.

openEuler is an open-source Linux distribution ecosystem, and its newest release ships focuses on running the whole cluster of AI machines as a single, unified system. The code is open and the community is showing interest in adopting it.

A Linux distro built for SuperPOD

openEuler 24.03 LTS SP3, released in December 2025, was the first openEuler version built for SuperPoD. openEuler calls it the world's first OS to support this architecture.

openEuler supports SuperPoD through UnifiedBus and its OS software stack. UnifiedBus provides the high-speed interconnect, while the UB OS Component and UB Service Core provide OS and cluster-level services on top of it.

A normal Linux system schedules memory and devices inside one node. UB lets applications reach memory and accelerators sitting in other nodes almost as if they were local. So this openEuler release breaks the single-machine boundary and schedules these distributed resources as one single set.

And as you can see, it matters most for the AI hardware. openEuler also did some of its own tests and claims that the peer-to-peer UB design boosts application performance in large-scale AI compute up to 30 to 50 percent.

Where SuperPoD OS comes from

Let me give you some context to why it's called "world's first SuperPoD OS".

openEuler grew out of Huawei's in-house EulerOS and is now an open-source distribution stewarded by the OpenAtom Foundation. It runs across Arm (including Huawei's Kunpeng), x86, and RISC-V, and the project has passed 16 million installations by the end of 2025.

SuperPoD and UnifiedBus are Huawei technologies used across multiple computing systems, including Atlas AI systems and the TaiShan 950 SuperPoD for general-purpose computing.

The UB OS Component is open sourced in the openEuler community and forms part of openEuler's software support for UnifiedBus.

Intelligence BooM, the full-stack AI layer

Another interesting thing that openEuler is pushing for is Intelligence BooM, its open-source, full-stack AI solution. Full-stack here means that instead of running as a tool, it covers the whole software layer an AI workload needs from the OS up.

Basically, it gives you an out-of-the-box setup for training and running large models, so users don't need to do all the "AI plumbing" from scratch.

Underneath, openEuler keeps working on the low-level plumbing that makes clusters efficient: combining different chip types into one usable pool (heterogeneous compute fusion), letting CPUs and NPUs work on the same job together (co-computing), and sharing memory across chips and machines (memory pooling).

It also splits a single accelerator so several jobs can share it and packs more than one workload onto the same hardware, both aimed at keeping expensive NPUs and XPUs busy.

For regular Joes like you and me, the most interesting part is the DevStation, openEuler's desktop build, which ships with an AI assistant that should handle things like installing software or setting up an environment from natural language commands.

We plan to test DevStation and its openEuler Intelligence assistant on an openEuler 24.03 test machine for day-to-day development and general-purpose use.

Moving towards a community ecosystem

It is interesting to see that most of the open source and open-weight AI models are coming from China.

From the enterprise focused UB Service core to user-focused Intelligence Boom, openEuler is keeping things open and that's a step in the right direction. It is certainly a project to keep an eye on.

If you want to get involved, check out the community page or their GitHub repo for more details.

So far, Canonical has followed a split kernel SRU cycle for Ubuntu, where regular fixes and security patches lived on separate tracks, with a full update every four weeks and a security-focused release at the two-week midpoint for urgent CVE fixes.

If that concept is new to you, a Stable Release Update, or SRU, is how Canonical ships bug fixes and security patches to Ubuntu after a release is out. The kernel gets its own dedicated track for this.

Canonical is replacing both tracks with a single 2-week cycle, and since each new cycle kicks off a week into the current one, this results in a kernel release landing every week.

The new 2-week cycle

Each cycle starts with a week of patch integration and prep work. The kernel team selects which fixes land on each kernel, builds the packages, and runs basic smoke tests to catch anything obviously wrong before the build moves on.

Those builds get pushed into Ubuntu's -proposed pocket once the first week wraps up. That is where kernel release candidates live before they have been certified, accessible to those who know where to look but not yet out to general users.

Week two is where Canonical runs the builds through its Ubuntu Certified hardware testing program, putting them through different machine types to make sure nothing breaks in the real world before the kernel ships.

A fresh cycle starts every week regardless of where the current one stands, so there is always a kernel finishing its test run and rolling out. That is how a 2-week cycle ends up delivering a release every week.

And, when something is not safe to ship, Canonical says that they will be clear about that and direct users toward general system hardening advice.

For teams that cannot wait the full two weeks, Canonical is explicitly pointing to -proposed as a fast path. The idea is that you run your own acceptance tests on the build sitting there rather than waiting for certification to wrap up.

Clankers made this inevitable

This change did not come from nowhere; they had to take such a sweeping decision due to clankers. Only last month, we saw how they were bleeding compute resources from git.kernel.org just by scraping it for training data.

For Canonical, their decision was driven by LLMs and AI agents turning vulnerability hunting into something automated and relentless, finding kernel bugs at a scale and speed no individual human researcher could replicate.

The goal is to have a workaround published within 24 to 48 hours of a CVE going public. Not the patch, but something that gets affected systems into a safer state while one is being built.

What the Ubuntu maker is doing here is responding to a rapidly-evolving situation by shortening the window between a CVE going public and a patch landing.

First thing first. Yesterday, I mistakenly sent the Local AI Weekly newsletter to everyone. That was not supposed to happen. I'll be more careful in the future.

Let's go with our regular dose of Linux and open source.

Snapdragon X2 processor support should be coming to Linux soon. Qualcomm has put out an early developer preview with Debian 13 as the reference environment.

PeppermintOS is switching from Xorg to XLibre for its upcoming Debian and Devuan ISO refreshes, with no plans to move to Wayland for the foreseeable future.

Mozilla and Mistral announced a partnership that adds Mistral Small 4 to Firefox Smart Window, swapping out the OpenAI model that was there before. Both of them believe that an independent browser and a European AI lab act as counterweights in the Big Tech-dominated browser ecosystem.

Valve has quietly open-sourced Lepton, the Android compatibility layer it built for Steam Frame. The tool runs Android games inside a container using Waydroid and LineageOS.

Garuda Linux has graduated its NixOS side project into a proper standalone distro. Garuda Nix now ships bootable ISOs for the Dr460nized and Mokka flavors.

After a decade of the open source Community Edition, Portainer is now changing things. CE stays frozen on 2.x with security patches and some backports, while Portainer 3.0, a Kubernetes-first rebuild, goes closed source.

GrapheneOS is publicly calling out Google for holding Android 17 QPR1 security patches back from non-Pixel OEMs until December.

After years of requiring a phone number even after adding username support, Signal is finally testing numberless registrations.

Then there are the two Ubuntu-related news items this week:

• Ubuntu 26.10 is changing how the desktop handles low memory, giving session-critical services like GNOME Shell a lower OOM score so your browser gets killed before your desktop does.
• Separately, the Ubuntu Wiki got its first comprehensive overhaul since 2010, now running on MediaWiki with niceties like dark mode, mobile support, and a WYSIWYG editor.

This edition of FOSS Weekly is supported by ANY.RUN that is organizing a webinar.

🧠 What We’re Thinking About

Andy Nguyen, lead of the PS5 Linux project, is stepping away after a hypervisor exploit he'd been sitting on got reported to Sony's bug bounty program by someone else before he was ready.

🧮 Linux Tips, Tutorials, and Learnings

Most KDE users open Konsole, type commands, and never bother tweaking it. We have some useful tips for you that will change how you use it.

Linux has had NTFS support for years, and it's never been great. NTFS-3G works but is slow and CPU-hungry. The ntfs3 driver is in the kernel but barely maintained. NTFSPlus is the attempt to finally do it properly.

Something broke on your Linux system, and you don't know where to start? Turns out systemd has been keeping detailed records the whole time. You just need to learn how to make good use of the tools that can help you read them.

👷 AI, Homelab and Hardware Corner

I got my hands on Elecrow's new HMI display, a 10.1-inch ESP32-P4 touchscreen built for homelab and maker projects, not for people who want something that works out of the box.

The recently released openKylin 3.0 comes with built-in skills that allow you to control and change system settings via AI agents.

✨ Apps and Projects Highlights

I did a fun project last weekend. I made a Wordle-like puzzle game with Linux terms. It's called Tuxdle.

Every day, you get a new word, and you have to guess the Linux term within 6 attempts. The term could be a distro, a command or popular terms like grub. Rules are explained here. Give it a try and let me know of your feedback, please.

Tasks.org is an open source Android to-do and reminders app that doesn't bombard you with ads nor does it trick you into a data collection scheme.

📽️ Videos for You

Make music listening more enjoyable with this GNOME extension? 👇

💡 Quick Handy Tip

Most of the time, you really don't need to install a fully-fledged theme to customize KDE Plasma. You can make do by installing a color scheme of your liking.

First, you need to download the .colors file for the theme of your liking. Then, inside the settings menu, go to Colors & Themes -> Colors. Here, click on "Install from file..." and select the theme file you want to install.

Ensure that the dropdown menu near the accent color palettes says "Accent color from color scheme," then click on the newly added theme, then on "Apply" to get it done.

🎋 Fun in the FOSSverse

How well-versed are you with the jargon of today's AI age? We are getting back to posting puzzles regularly, so give this new one a try!

Linux Mint is one of the OG Linux projects that deserves more funding, you know. 🫠

🗓️ Tech Trivia: The first successful FORTRAN program ran at IBM on September 20, 1954, pioneering a way for programmers to express computations using algebra-like notation rather than machine code.

The programming language was formally released in 1957 and remains in use today, particularly in science and engineering.

🧑‍🤝‍🧑 From the Community: Daniel was facing an issue with YouTube playing back videos with a delay. He found an open source tool that did the job without much fuss.

Qualcomm has put out an early developer preview of Linux for Snapdragon X2 Series laptops, asking kernel developers, distro maintainers, and hardware enablement engineers to get on this hardware and start testing.

They have been laying the groundwork for this since Snapdragon Summit 2025, and the push today is toward getting drivers into mainline rather than shipping out-of-tree patches that only a handful of people can maintain.

This is not aimed at end users.

The progress so far

The current preview pairs a Debian 13 "Trixie" userspace with a custom kernel, and the initial focus is on readying up the core pieces developers need before they can do any serious work on the hardware.

System boots are handled via systemd-boot acting as the UEFI manager, while Core I/O covers USB, PCIe, and the Qualcomm Universal Peripheral serial engine for UART, I2C, and SPI connections.

Whereas Mesa's open source Qualcomm drivers handle graphics duties, with Freedreno covering OpenGL, Turnip taking care of Vulkan 1.3, and Rusticl handling OpenCL compute.

Work on all three is ongoing, but display output, browser acceleration, and GPU compute workloads are already possible.

The FastRPC driver, Qualcomm's mechanism for routing compute tasks from the CPU to the DSP, is being pushed upstream to open up the Hexagon NPU for local AI inference workloads. Power management and thermal support are part of this series of developments too.

Two build paths

The official documentation for Snapdragon X2 Linux software outlines two paths that developers can follow.

The first is the unregistered path, where you build the Debian OS layer using prebuilt firmware binaries, getting started without a Qualcomm developer account. The second requires an account, giving you full firmware source access to build everything from scratch.

Prebuilt flashable images are also available if you want to skip the build entirely and go straight to flashing.

Host requirements also differ by what you are building. The Debian layer needs an ARM64 machine running Ubuntu 24.04 LTS or later, or Debian Trixie. The firmware build requires an x86_64 host on Ubuntu 22.04 LTS or later.

If you want to see a hands-on demo of Linux unning on Snapdragon X2 hardware, then this year's Snapdragon Summit (ends today) in Maui, Hawaii, is the place to be!

More of this, please

A staged roadmap, a documentation portal, dedicated build scripts in a public repo, and turning to developers for feedback is something other silicon vendors looking to upstream their work should take note of.

In the end, Linux wins. Every driver that lands in mainline is one fewer patch that distributions have to carry, and that is what eventually makes these machines usable beyond early adopters.

Suggested Read 📖: A Linux distro that doesn't plan to stay around long.

openKylin is China's homegrown, open source operating system which is developed by a consortium of companies, research institutes, and individual contributors.

We have covered openKylin in the past but that was more than two years ago. There is a new major release in the form of openKylin 3.0 that brings Linux kernel 7.0, UKUI (desktop environment) 4.24 and deeply integrated AI at the core.

Let's see the highlight of the new release.

AI is at the center of openKylin 3.0

Among the most highlighted changes in the new release is the built-in, all-new KylinBot AI agent framework, along with Xiao K, the AI assistant built on top of it.

Xiao K is basically the graphical interface similar to Hermes Desktop and most other AI tools of this sort. The underlying KylineBot AI agent framework integrates it with the system.

And then it comes with several pre-built system skills like:

• Input Devices: Full coverage for keyboard, mouse, and touchpad.
• Display Control: Screen brightness and OSD display adjustments.
• Network Management: Deep integration with the NetworkManager command-line tool to manage Wi-Fi, Ethernet, and VPNs with a single command.
• Peripheral Management: Bluetooth toggle, device discovery, and automatic audio device pairing.
• System Controls: Power management, taskbar operations, shortcut navigation, and system app controls

This way, you can ask the agent in Xiao K to switch off the bluetooth or switch to a different network, change the input keyboard layout and more. You don't need to go through the command line or click through the system settings. The agent does the job for you.

Seems like overkill, right? But it makes more sense if you use the voice input. People these days use Siri hands-off for calling to specific contacts. So, if you could just talk to your computer and ask it to connect to your bluetooth headphone, why not?

Remember that Ubuntu 26.10 is also working on AI integration. Omarch 4.0 got a sudden surge in popularity for AI (and its looks). So, even if you dislike AI, there is a demand for it. And sooner or later, most desktop operating systems will have built-in AI assistants like this. openKylin is just a few steps ahead of its peers.

There are more skills that can be found on openKylin's repo here. It is in Chinese language and I do hope they start offering these features in English as well to cater a larger global audience.

Like Hermes, OpenClaw and other AI tools, Xiao K doesn't stay to your desktop only. You can communicate with it from messaging tools like WeChat, Feishu, DingTalk etc. Telegram and Discord support is planned.

UKUI version 4.24

The UKUI (Ultimate Kylin User Interface) desktop environment is upgraded to version 4.24. The new version brings the following updates:

Standardized API encapsulation across UKUI modules and integration with KACP (Kylin Agent Control Protocol), enabling KylinBot to invoke underlying desktop capabilities via unified APIs.

The new version also brings accessibility support in the form of a screen reader. It also has air gesture controls, which means you can control some desktop functionality with the movement of the hands instead of touching the screen or using a mouse or keyboard.

This works on Android devices thanks to the proximity sensor in the smartphone. Here, Kylin uses the webcam to capture the hand gestures.

The feature to enable dynamic wallpaper is also added for enhanced eye candy. In other visual changes, "Three-Island" taskbar layout mode is added as an option.

Touchpad gets enhanced multi-finger gesture interactions, supporting multitask view, volume control, and other actions using 3-finger and 4-finger gestures.

Rustification

Since both AI and Rust are in trend these days, openKylin 3 has new Rust-based utilities that replace the older, classic utilities.

kylin-time replaces the legacy time utility, kylin-wget replaces the classic wget and kylin-ki18n replaces the ki18n.

Getting openKylin

If you want to give it a try, you can download openKylin from the official website. The ISO is a little over 7 GB. The size is primarily because the OS comes with pre-installed AI models.

I plan to test openKylin 3.0 on my spare system. It is okay in VM, but since there is a lot of AI integration and I would like to use them to see how effectively they work on a baremetal system. I will share my findings in a review video. Stay tuned for that.