AI

Linux-AI OS Review: LMDE 7 With a Built-In Ollama Assistant

Does a distro that ships Ollama and two Llama models inside its ISO save you real work? That’s what I wanted this Linux-AI OS review to settle. I downloaded the free 1.0 “Star” image, checked it, installed it in a UEFI virtual machine and used the built-in assistant until it broke. It broke in three distinct ways, and the install left a handful of messes behind that you’ll want to clean up.

Original content from computingforgeeks.com - post 182989

You’ll get the measured ISO size and checksum, the installed footprint and idle RAM, and what the assistant really is. I also timed its models on CPU, checked what the published root password unlocks, and built the same thing on stock LMDE 7’s base in three commands. Premium is not covered beyond the vendor’s own page, because I only tested the free edition. I ran the whole review in November 2026 on Linux-AI OS 1.0 Star in a UEFI VM with 6 vCPUs, 12 GB of RAM and no GPU, and every screen and number below comes from that box.

Sorting out what Linux-AI OS actually is

Linux-AI OS is a Cinnamon desktop on LMDE 7 “Gigi”, which in turn sits on Debian 13 “Trixie”. The installed system pulls packages from the Debian trixie and trixie-security repositories plus Mint’s gigi repository, and /etc/linuxmint/info still carries Mint’s file layout with the branding swapped to “Linux-AI”. What the project adds is a GTK assistant, a Welcome Wizard, a Control Center, an app installer, desktop style presets, and Ollama with models already on disk.

The project is anonymous: the site names no team, and the GitHub repository holds only a README with no license file. I did not expect what I found inside the ISO, though. It is not built with Mint’s own live tooling. The image boots a Linux Live Kit style layout (initrfs.img, a bootinst.sh script and one 13 GB 01-core.sb squashfs), and the installer is Calamares rather than LMDE’s live installer. That matters later, because the leftovers I clean up after the install come from how this Calamares setup finishes the job.

The free edition and the paid one split like this, according to the vendor’s Premium page (I didn’t buy or test Premium):

ItemStar (free)Premium StandardPremium AI Studio
PriceFree download$49.95 one-time, access to both editions, you pick one ISO to downloadSame purchase
Cinnamon AI desktop and assistantIncludedIncludedIncluded
Mail assistant, voice studio, image and video tools, VM manager, audio recorderNot includedIncludedIncluded
Music Studio, PersonaStudioNot includedNot includedIncluded
Vendor GPU guidanceNone for the base desktopNot statedMusic Studio: 8 GB VRAM minimum, 12 GB or more recommended

Keep the “Not included” rows in mind, because the free ISO I downloaded ships the code for several of those Premium apps anyway. I cover that further down.

Downloading and checking the Linux-AI OS ISO

My Linux-AI OS download came to 13,740,867,584 bytes, which is the “12.80 GB” the vendor states once you read it as GiB, and its SHA-256 matched the published sha256sum.txt on November 2, 2026. The vendor’s homepage and download page both say 12.80 GB today, and no vendor page I read lists the 11.61 GB figure that circulated earlier. In September, pillitteri.it found a 14,119,665,664-byte image dated September 16 and a checksum file that named a different ISO (linux-ai-star-v16.iso). Neither applies now: the image on the server is an October 8 build, and the checksum file names the file you actually get.

Put the ISO and the checksum file in the same directory before you check it, because sha256sum -c reads the file name from the checksum file and looks for it next to you:

sha256sum -c sha256sum.txt

A good download prints the file name followed by OK. Anything else means a truncated or altered image, and at 12.8 GB a truncated download is the likelier of the two:

Linux-AI OS ISO size of 13740867584 bytes and sha256sum check returning OK

My first copy was the slow lesson here. The vendor’s server gave a single stream as little as 26 KB/s into my lab, so I finished the download with eight parallel curl -r range requests and let the checksum prove the parts went back together correctly. If you hit the same crawl, the download page also offers the ISO in smaller parts with rebuild instructions.

Two details on the image itself decide how you can boot it:

  • GRUB boots it, and the vendor only promises “Compatible with X64 UEFI”
  • the leftover syslinux config points at /Linux-AI-Star/Linux-AI-Star/boot/, a path that doesn’t exist
  • the kernel command line passes security=none apparmor=0, so AppArmor is off in the live session

I only booted it under UEFI (OVMF), so I cannot tell you what a legacy BIOS machine does with it. The ISO’s BIOS boot record points at GRUB too, and the broken syslinux path only matters if you use its bootinst.sh script, but with the vendor promising UEFI alone I’d stick to UEFI. USB writers such as Etcher and Ventoy are also outside what I tested, because a VM boots the ISO file directly.

Installing Linux-AI OS in a UEFI VM

The live desktop came up within a minute with the Welcome Wizard already open, and the “Install Linux-AI Star” icon starts Calamares as root without asking for a password. The live user, linux-ai, has passwordless sudo, which is normal for a live session. The installer asks the usual six questions (language, location, keyboard, partitions, user, summary), and the partition page is the one worth reading before you click Next:

Linux-AI OS Calamares Erase disk option planning a 12.85 GiB swap partition
  1. 1Erase disk
  2. 212.85 GiB swap taken from the root disk
  1. Erase disk
  2. 12.85 GiB swap taken from the root disk

On an 80 GB disk with 12 GB of RAM, Erase disk plans a 300 MiB EFI partition, a 66.86 GiB ext4 root and a swap partition the size of your RAM. That’s a fair default for hibernation, but on a 32 GB workstation it would carve 32 GB off a 120 GB SSD, so pick Manual partitioning if your RAM is large and your disk isn’t. There’s no encryption by default either; the “Encrypt system” box is unticked.

The Users page is where the install first went wrong for me:

Linux-AI OS Calamares Users page with no root password field
  1. Computer name the installer ignores

I typed box1 as the computer name, and the installed system still booted as linux-ai, with /etc/hosts saying the same. I couldn’t find which install step overwrites it, because Calamares’ own users module normally writes that file, so treat it as a known quirk and fix it after the first boot.

As for the missing root password field, it is deliberate: /etc/calamares/modules/users.conf sets setRootPassword: false, which in Calamares means the root account stays disabled. The copy and setup took about 20 minutes on this VM (28% at the 12.7-minute mark), most of it spent unpacking the xz-compressed squashfs on the CPU.

The finish page has “Restart now” pre-ticked, so clicking Done reboots straight away. Pull the USB stick or detach the ISO first, or you land back in the live session, which is exactly what happened to me. After that, GRUB shows a single “Linux-AI Star” entry and LightDM greets you with the account you created:

Linux-AI OS first boot with the Welcome Wizard and Cinnamon desktop

The Welcome Wizard starts at login from an autostart entry (linux-ai-firstboot-welcome.desktop), and the desktop carries icons for the assistant, the Control Center, the App Installer and an Audio Recorder. That last one is a Premium app, which I’ll come back to.

Measuring the installed system

Linux-AI OS idles at about 1.2 GB of RAM and uses 22 GB of disk right after install, and roughly 10 GB of that is AI payload you may never use. I rebooted, logged in on the desktop, and read memory, disk and Ollama’s state over SSH three minutes after boot. One caveat on the SSH part: the distro doesn’t ship an SSH server, so I installed openssh-server myself for these captures.

Linux-AI OS idle RAM, root disk use and the two bundled Ollama models

1,226 MiB used is light for Cinnamon plus an always-on ollama serve, and it stays light because Ollama loads nothing until you ask a question (ollama ps is empty). The two models are llama3.2:1b (1.3 GB) and llama3:latest, the original 8B Llama 3 (4.7 GB), both pulled about five months before I installed. The breakdown of where the disk went is more interesting:

Linux-AI OS disk use by Ollama models, CUDA runtimes and the bundled apps

That 3.5 GB under /usr/local/lib/ollama is mostly CUDA 12 and CUDA 13 runtimes plus a Vulkan backend, which is Ollama’s standard Linux bundle rather than anything the distro tuned. The table puts the numbers that matter in one place:

What I measuredResult on Linux-AI OS 1.0 Star
ISO size13,740,867,584 bytes (12.80 GiB)
Root disk used after install22 GB of 66 GB
Idle RAM, 3 minutes after boot1,226 MiB used of 11,960 MiB
Ollama models on disk5.6 GB (llama3.2:1b and llama3)
Ollama runtime with CUDA and Vulkan libraries3.5 GB
Bundled Ollama version0.23.2 (install.sh gave 0.40.13 the same week)
Installed packages2,460
Pending upgrades on first boot396
Ollama API127.0.0.1:11434 only, enabled at boot

The Ollama version is the row I’d worry about. Ollama doesn’t come from an apt repository here; it’s a binary dropped into /usr/local/bin, so the 396 pending upgrades won’t touch it and nothing in the distro will. You update it the same way you’d install it on any distro, by rerunning the upstream install script.

Testing the Linux-AI assistant

The Linux-AI Assistant is a GTK 4 and libadwaita window written in Python (/usr/share/linux-ai/assistant/linux_ai_assistant.py) that sends each question to Ollama’s /api/generate endpoint with the same fixed instructions prepended to every question. The features page describes a Flask backend with an HTML/CSS frontend, but that isn’t what ships: there’s no web server, just the desktop app talking to Ollama on localhost. It also has no conversation memory, since every question goes out as a single prompt with no earlier turns attached.

Its model list and limits are hardcoded in that file, which the menu shows:

Linux-AI Assistant model menu with llama3.2:1b, llama3.2:3b and llama3
  1. 1llama3.2:3b is listed but not shipped
  2. 2Answer Length sets num_predict
  1. llama3.2:3b is listed but not shipped
  2. Answer Length sets num_predict

Only three models can ever be picked, whatever you pull with Ollama, because the menu is three append_text() calls. Answer Length maps to num_predict values of 180 (Short), 300 (Medium) and 550 (Detailed), and num_ctx is fixed at 2048 with temperature 0.3. You can’t change any of those from the window, only by editing the Python file. The assistant never sends a num_predict of 256, a value I’d seen attributed to it; its only values are 180, 300 and 550.

I asked the default model a question any Linux user eventually types, “How do I check which process is using the most memory on Linux?”, and the answer streamed in within a minute:

Linux-AI Assistant answer from llama3.2:1b with empty code fences and a broken sort command

Read it closely and it’s mostly wrong. The code fences are empty, and top -bn1 is described as “a buffer size of 1 second” when it means batch mode, one iteration. The suggested sort -rn -k2 -t | head -n 10 fails because -t needs a delimiter, and htop gets called graphical. The window also prints Markdown raw, so headings show up as ### and bold as **. Switching the menu to llama3 gave a correct answer (run top, press M to sort by memory) but took about two minutes.

To put numbers on that, I sent the same question through the API with exactly the options the assistant uses on Medium (num_predict 300, num_ctx 2048), with the models still warm from the GUI tests:

Ollama tokens per second on Linux-AI OS for llama3.2:1b, llama3.2:3b and llama3 on 6 vCPUs

On 6 vCPUs that’s 17.5 tokens per second for the 1B model, 10.9 for 3B and 5.2 for the 8B llama3. Treat those as the good case. A second run after a cold reboot came in at 15.8, 3.9 and 2.9, and I couldn’t pin down why the larger models dropped that far, so on shared or modest hardware expect the 8B model to crawl. Both larger models also hit the 300-token cap and stopped mid-answer in both runs, which is what Medium means in practice: switch to Detailed before you ask anything that needs a long reply. No model returned an empty reply in my tests. If you want a faster model on this kind of hardware, the CPU-only Ollama tests on Fedora compare more of them than the three this menu offers.

Will you hit the same problems? Probably, because all three come from the shipped configuration rather than my VM. Each one has the exact message as its heading.

Error: “404 Client Error: Not Found for url: http://localhost:11434/api/generate”

Pick llama3.2:3b from the menu and ask anything, and the window only says “Something went wrong”:

Linux-AI Assistant Something went wrong message after selecting llama3.2:3b

The detail is in ~/.local/share/linux-ai/assistant_errors.log, where the 404 comes from Ollama refusing a model it doesn’t have. The ISO lists 3B in the menu but never pulled it, so pull it once:

ollama pull llama3.2:3b

After the 2.0 GB download, ollama list shows three models and the 3B choice answers normally:

Linux-AI Assistant 404 Client Error in the log and the ollama pull llama3.2:3b fix

On speed alone, 3B is the middle ground: 10.9 tokens per second in my warm benchmark against 5.2 for the 8B llama3.

“Something went wrong” after an app install that worked

Typing “Install VLC” doesn’t touch the LLM at all. It matches a built-in catalogue of 48 apps by keyword, shows a confirmation dialog naming the package, then opens a terminal that runs sudo apt update; sudo apt install -y vlc and asks for your password. VLC installed fine, yet the assistant reported a failure:

Linux-AI Assistant reporting Something went wrong while the VLC install finishes in the terminal

The log says timed out after 60 seconds. The assistant waits for the terminal window with a 60-second limit, and that terminal stays open until you press Enter, so any install that takes longer than a minute gets flagged, including the time you spend typing your password. Trust the terminal output, not the assistant’s message. The live image even ships the builder’s own log from May 13, 2026 with the same timeout on “Install Chromium”, so this one has been known for a while.

“Whisper speech-to-text was not found”

Mic Ask records six seconds with arecord and hands the file to the whisper command with the base model. The free install has arecord and the Piper voices for spoken replies, but no whisper binary anywhere on the system, so the button stops here:

Linux-AI Assistant Mic Ask failing with Whisper speech-to-text was not found

So the voice input the release notes advertise does not work on a fresh free install. Spoken replies do, if your machine has audio. I haven’t tried installing a whisper CLI to fill the gap, so I can’t tell you which build the assistant expects.

Checking the Linux-AI OS root password

According to the download page, the Linux-AI OS root password is linux-ai, but on the installed system it doesn’t log you in as root at all. The claim sits in the page’s opening paragraph:

Linux-AI OS download page publishing the root password linux-ai
  1. Root password published on the download page

So what does that password actually open? In my live session it unlocked the screen lock of the live user, linux-ai, rather than a root login. That user has passwordless sudo, as the linux-ai-live sudoers rule shown below confirms, so the page’s wording is loose but the practical effect is the same in the live session. Anyone at the keyboard, or on the network if you add an SSH server, is root. Don’t leave the live session running on a network you don’t control.

On the installed system the story is better, with one leftover. I checked it from the terminal as my own user:

Linux-AI OS installed system with root locked, su failing with linux-ai and the leftover NOPASSWD sudoers rule

Root shows L (locked), so su - with linux-ai fails with “Authentication failure”, and the published password is useless against an installed machine. The leftover is /etc/sudoers.d/linux-ai-live, which still grants linux-ai ALL=(ALL) NOPASSWD: ALL. The live account itself is gone after install (getent passwd linux-ai returns nothing, passwd -S linux-ai says the user doesn’t exist, and /home holds only my account), so the rule does nothing today, but it would hand passwordless root to anyone who later creates that user name. Delete it, then confirm the remaining sudoers files still parse:

sudo rm /etc/sudoers.d/linux-ai-live
sudo visudo -c

On my box visudo -c listed /etc/sudoers and the four remaining files under /etc/sudoers.d as “parsed OK”. One of those, 10-installer, holds the ordinary %sudo ALL=(ALL) ALL rule, so your own account keeps sudo with a password, which is what you want.

Cleaning up apt, the hostname and ufw

Every apt update on a fresh install ends in an error, and the cause is a live-media source the installer leaves behind. This is the message you’ll see, from the Update Manager or the terminal:

E: Failed to fetch file:/dists/trixie/main/binary-amd64/Packages  File not found - /dists/trixie/main/binary-amd64/Packages (2: No such file or directory)
E: Some index files failed to download. They have been ignored, or old ones used instead.

The file is /etc/apt/sources.list.d/debian-live-media.list, holding deb [trusted=yes] file: trixie main. Besides the error, trusted=yes tells apt to skip signature checks for that source, which is a bad line to keep around even when it points nowhere. Removing it fixes both:

sudo rm /etc/apt/sources.list.d/debian-live-media.list
sudo apt update

The next run reads only the Debian and Mint repositories and finishes without an E: line (the capture shows the tail of it):

Removing debian-live-media.list so apt update runs clean on Linux-AI OS

While you’re in a terminal, set the hostname the installer ignored. hostnamectl changes it immediately, and the sed line keeps /etc/hosts in step so sudo doesn’t complain about resolving the old name:

sudo hostnamectl set-hostname box1
sudo sed -i "s/\blinux-ai\b/box1/" /etc/hosts

Running hostnamectl hostname afterwards printed box1, and /etc/hosts showed 127.0.1.1 box1. Swap in your own name, and check /etc/hosts afterwards, because the sed pattern only matches the exact word linux-ai.

The firewall is the third leftover. ufw printed ERROR: Couldn't stat '/etc/ufw/user.rules' when I installed openssh-server, and the image ships none of ufw’s six rule files in /etc/ufw. Reinstalling the package only put the two user*.rules files back, because dpkg won’t replace config files an admin deleted, and ufw enable then stopped at ERROR: Couldn't stat '/etc/ufw/before.rules'. On an untouched install expect the same message for user.rules first.

Luckily the templates are still on disk, so copy them in, allow SSH if you run a server, and turn the firewall on:

sudo cp /usr/share/ufw/iptables/before.rules /usr/share/ufw/iptables/before6.rules /usr/share/ufw/iptables/after.rules /usr/share/ufw/iptables/after6.rules /usr/share/ufw/iptables/user.rules /usr/share/ufw/iptables/user6.rules /etc/ufw/
sudo chmod 640 /etc/ufw/*.rules
sudo ufw allow ssh
sudo ufw enable

After that, sudo ufw enable answered “Firewall is active and enabled on system startup”, and sudo ufw status verbose showed deny incoming, allow outgoing and the 22/tcp rule. My scripted run used ufw --force enable; run interactively over SSH, plain ufw enable asks whether to proceed, and answering y is safe because SSH is already allowed. Skip the allow ssh line on a desktop with no SSH server. The rest of the Linux Mint post-install checklist applies to this Debian-based desktop too, once the firewall works.

Finding Premium apps on the free ISO

The free ISO carries the Premium application folders under /opt/linux-ai/apps, 987 MB of them, and the Audio Recorder from that set sits on the desktop and runs. The vendor’s comparison table lists the “Professional Audio Recorder” as Premium only, yet double-clicking the icon on my free install opened it straight away:

Linux-AI Audio Recorder running on the free Linux-AI OS install

The folders that ship on the free image are:

  • audio-recorder, launched from the desktop icon
  • mail-assistant, ollama-manager, vm-manager
  • image-maker and image-video-maker, with about 400 MB of overlay videos
  • voice-studio, with a Kokoro TTS model and its own Python venv

Only the Audio Recorder has a desktop or menu entry. A linux-ai-premium launcher script also ships in /usr/local/bin, but I did not try to run it or the other apps, since they’re what Premium buyers pay for. I read this as packaging leakage rather than a hidden bonus: the free and Premium builds clearly come from one tree, and nobody pruned it before the free release. If you’re weighing the $49.95, know that I can’t say how the Premium builds behave, because I never had one.

Building the same setup on LMDE 7’s Debian base or Mint 22

You can run an Ollama assistant on LMDE 7’s Debian base or on Mint 22 with three commands and get a newer Ollama than Linux-AI ships. My Linux Mint 22.3 test VM sat on a Proxmox node that kept rebooting during this job, so the timed run below is on stock Debian 13, the base LMDE 7 is built on. Mint 22 is Ubuntu 24.04 underneath, so any Ollama Ubuntu 24.04 walkthrough applies to it, and our guide to installing Ollama on Rocky Linux 10 / Ubuntu 24.04 covers that base step by step.

  1. Install Ollama with the upstream script
  2. Pull the same default model Linux-AI uses
  3. Add a desktop chat window from Flathub

Those three steps are these commands, in order:

curl -fsSL https://ollama.com/install.sh | sh
ollama pull llama3.2:1b
sudo flatpak install -y flathub com.jeffser.Alpaca

Mint and LMDE come with Flatpak and Flathub already set up. My Debian 13 box didn’t, so I added the flatpak package and the Flathub remote first, which took another minute. Each step’s timing on my lab’s connection that evening:

Timed install of Ollama, llama3.2:1b and Alpaca on stock Debian 13

38 minutes of wall clock sounds slow, but nearly all of it was download time on a home line that was crawling that night. The Ollama step alone spent 27 minutes pulling a 1.5 GB tarball at under 1 MB/s, while the 1.3 GB model came down at about 2.6 MB/s. The hands-on part is three commands, and a faster line would shorten the wait, but I didn’t time a run on Mint or LMDE, so budget for the downloads rather than a fixed number.

What you end up with is Ollama 0.40.13 instead of 0.23.2, a model you chose, and Alpaca, an Ollama chat client that keeps conversations and lets you pull and switch models. Budget the disk: Flatpak reports 2.3 GB for the app, and /var/lib/flatpak reached 4.6 GB once the GNOME runtimes came along. I installed Alpaca on that headless test box but did not open its window there, so I haven’t verified how it finds the local Ollama server; the Ollama API it needs listens on 127.0.0.1:11434.

If you’d rather have a browser UI, the Open WebUI setup for Ollama runs on the same box, and the Ubuntu 26.04 Ollama and Open WebUI guide walks through both together. For picking something better than a 1B model, the Ollama models cheat sheet lists sizes and pull names. And if you plan to run 8B models daily, a GPU changes the picture far more than any distro does; our RTX 5070 Ti vs 5060 Ti local AI comparison has real token rates.

Deciding whether you need Linux-AI OS

For most readers, no, at least going by the Debian 13 base I tested. There, three commands and 38 minutes of mostly download time gave me a newer Ollama and a model of my choice without the apt, sudoers and ufw leftovers above. Alpaca installed cleanly there, but I never opened its window or verified its connection to Ollama, so I can’t yet vouch for it as a replacement for the assistant. Linux-AI OS earns its place in one narrow case, an offline machine. Because the models ship inside the ISO, it’s the only setup here where the assistant answers on first boot with the network cable unplugged. That’s handy for a classroom demo, or for a box used by someone who’ll never open a terminal.

Here’s how I’d decide:

  • already on Mint or LMDE: add Ollama and skip this distro (tested on the Debian 13 base; Alpaca’s Ollama link unverified)
  • need AI on an offline machine: Linux-AI OS, then pull llama3.2:3b while you’re still online
  • handing a desktop to a non-technical relative: possible, after the sudoers, apt, hostname and ufw cleanups
  • running anything you can’t afford to lose: a mainstream distro with a known maintainer

The anonymous project and the missing license weigh more than the bugs for that last group. The code is readable Python on disk, which is more than many “AI distro” projects offer, but nobody has said who maintains it or under what terms. If you’re still choosing a base desktop, our best Linux desktop distributions roundup is a better starting point.

What I still don’t know

A few things I couldn’t settle in the lab, and I’d like answers before recommending this to anyone:

  • whether a legacy BIOS PC boots and installs it, since only UEFI is promised
  • how Ventoy and Etcher handle a 12.8 GB image on real USB sticks
  • which whisper build Mic Ask expects, and whether Premium installs it
  • whether updates will ever move the bundled Ollama off 0.23.2

The Ollama question is the one I’d watch. The assistant, the models and the CUDA libraries all hang off a binary that the package manager cannot see. If the project goes quiet, every install stays on that version until its owner reruns the upstream script by hand.

Keep reading

Ollama Models Cheat Sheet: Best Models by VRAM, Tested (PDF) AI Ollama Models Cheat Sheet: Best Models by VRAM, Tested (PDF) Open Source LLM Comparison Table (2026) AI Open Source LLM Comparison Table (2026) Claude Code Cheat Sheet: Commands, Shortcuts and Flags (PDF) AI Claude Code Cheat Sheet: Commands, Shortcuts and Flags (PDF) Dual GPU LLM Build: PCIe x8/x8 vs x16/x4, PSU and Cooling AI Dual GPU LLM Build: PCIe x8/x8 vs x16/x4, PSU and Cooling MCP vs RAG vs Agent Skills: When to Use Each AI MCP vs RAG vs Agent Skills: When to Use Each Aptakube Kubernetes GUI: Install and Multi-Cluster Setup Containers Aptakube Kubernetes GUI: Install and Multi-Cluster Setup

Leave a Comment

Press ESC to close