AI-2: Linux With an AI Brain

In this post

• Meet AI-2

• Inside the Toolkit

• Getting Started

Most conversations about artificial intelligence assume a shiny new laptop, a hefty graphics card or a monthly subscription. AI-2 politely disagrees. It is a Linux distribution built on the idea that the tired old PC in your cupboard deserves an AI assistant too: one that lives entirely on the machine, asks for no account and sends nothing anywhere. It is young, refreshingly candid about its limits and rather charming. Let us take a proper look.

Meet AI-2

AI-2 is a Linux system for ordinary and older computers that gives them a local AI assistant, a language model running on the machine itself, on the CPU, with no account, no subscription and no data leaving the computer. Its tagline says it all: give your computer an AI brain.

Under the bonnet sits Artix Linux, the Arch-derived system that swaps systemd for runit, paired with the Xfce desktop and a rolling-release model.

You install once and update ever after with pacman, so there is no reinstalling for a new version. DistroWatch files AI-2 under Arch and Artix, lists its origin as Germany and Spain, and places it in the Desktop, Large Language Model and Live Medium categories. It is made by Rafael Minuesa of ProWoos and licensed under the MIT licence, and its repository notes that it was built with Claude Code. Artix Linux, llama.cpp and the models themselves keep their own licences.

The project is brand new. Its SourceForge page was registered on 23 August 2026, and the GitHub releases page now lists a steady run of dated ISO builds. The latest, ISO 20261001, arrived on 1 October 2026 and carries AI-2 0.19.3. It weighs 2,062,987,264 bytes, a touch over 2 GB in decimal terms and slightly heavier than the 1.85 GB quoted in older project text. For that build DistroWatch tracks Linux 7.2.7, glibc 2.44 and Xfce's xfdesktop 4.20.2, and lists Epiphany as the default browser.

Development has been brisk. ISO 20260917 (AI-2 0.18.0) added the Search Knowledge menu entry and the w3m text browser. The next day's build, 0.18.6, taught that search to open a whole document at the right passage, and to offer installing the packs on a machine that lacks them. ISO 20260921 (0.19.0) turned the three Knowledge Packs into proper packages, so a routine pacman update refreshes them, and began signing the ISO checksum. Version 0.19.1 tidied the setup wizard's Knowledge Packs screen, 0.19.2 taught the installer Polish thanks to Mateusz Szczepaniak, and 0.19.3 gave the first-login window a friendly "setup is starting, please wait" message while silencing an update notification that cried wolf after a fresh update.

Early images had teething troubles, and the FAQ documents them candidly. ISOs 20260816 and 20260817 lost Artix's default GRUB settings, so an install booted to a plain text menu without Windows listed; images from 20260819 onwards are fixed, and the FAQ gives the repair for affected systems. Another early quirk hid the "Erase disk" option on machines that had previously run Linux, fixed in ISO 20260825.

The requirements are gentle. You need a 64-bit x86_64 PC, 2 GB of RAM (4 GB recommended) and about 6 GB of disk. BIOS/MBR and UEFI/GPT installs both work, although Secure Boot must be switched off because the image is not signed for it. Thirty-two-bit machines are not supported, and the boot menu says so plainly. Above that floor, AI-2 does not turn a computer away for being old or slow, and CPUs without AVX or even SSE4.1 are served by dedicated engine builds.

What truly sets AI-2 apart is its honesty. The project's recurring theme is that RAM tells you what fits, not what runs. So rather than promising a grand model that crawls, AI-2 measures your hardware and tells you the truth. If a local model would be unusably slow, it suggests remote inference for the heavy work, but only by explicit choice. Local where possible, remote only if you say so.

That philosophy was forged on real hardware. The oldest validated machine is a 2011 HP Pavilion g4 with an AMD A4-3305M, 4 GB of RAM and a spinning disk. It lands in the Light tier with an AI Score of 29 to 30. There, Gemma 3 270M manages about 4.5 tokens per second, Qwen2.5 0.5B around 1.8 to 2.0, and browser chat roughly 1.45. That is patience mode, as the wiki admits, but genuinely local and genuinely working. A BIOS/MBR install alongside Windows 7 has been verified on it too.

Then comes a delightful plot twist. A 2016 Acer ES1-522 sits in the Standard tier yet scores just 23, below the 2011 laptop, proving that newer is not automatically faster. Installing on it also produced two real-world fixes: older AMD graphics from the 2012 to 2016 era are now steered to the mature radeon driver, because amdgpu crashed on them and left the screen black, and leftover swap partitions are released before the installer runs.

Beyond those two laptops, every ISO is verified by complete QEMU installs in both BIOS and UEFI modes, including first login, the setup wizard and chat. "Validated" has a clear meaning here: the release checklist covers a full installer run, a branded login, a wizard that completes, tuning that survives a reboot and chat that answers in the browser, not merely a machine that booted once. The project calls itself young and says so openly: tested on two older AMD laptops and in virtual machines, so yours may well be the third.

Inside the Toolkit

The magic begins at first login. A setup window opens by itself and, in a few minutes and three questions, does the heavy lifting. It scans the computer, assigns a capability tier, tunes the system, installs the right engine build, measures an AI Score, recommends a model and offers to download it. It even finishes without a network, noting that the model and score are pending until you are online. Missed it? Run ai-2 wizard whenever you like.

There are six tiers: Tiny (2 GB of RAM, one core), Light (4 GB, two cores), Standard (8 GB, two cores), Creator (16 GB, four cores), Studio (32 GB, four cores) and Workstation (64 GB, eight cores). A machine gets the highest tier whose requirements it meets, and one below every requirement still receives the smallest tier. The tier shapes the tuning. Compressed swap in RAM uses zram on the small tiers and zswap from Standard upwards, earlyoom guards against running out of memory on every tier, and idle suspend is disabled so a long download is never cut short. On Tiny and Light the AI engine starts on demand and exits after five or ten idle minutes, handing the RAM back; from Standard up it runs as a persistent service. The smallest tier also keeps services an old machine does not need, such as cupsd, avahi-daemon and bluetoothd, switched off, and it has no voice input.

The AI Score is the cleverest idea here. ai-2 benchmark always runs the same fixed workload, Qwen2.5 0.5B Instruct in Q4_K_M form under llama.cpp's llama-bench, so scores stay comparable across machines. AI-2 refuses to benchmark on a substitute model. The run is time-boxed to between 150 and 300 seconds, and generation speed is mapped to a 0 to 100 scale: about 2 tokens per second scores roughly 30, around 10 scores roughly 65 and feels comfortable for chat, and 40 or more earns the full 100. Alongside the number come star ratings from 0 to 5 for eight capabilities: chat, translation, programming, OCR, document Q&A, voice, image generation and video. The last two sit at zero stars on every machine today, because the packaged engine is CPU-only.

That CPU-only stance is deliberate. A review dated 14 September 2026 concluded that CUDA and ROCm are not coming for the hardware AI-2 targets: the CUDA toolkit package alone exceeds the whole ISO in size, and ROCm weighs over 9 GB installed. Proprietary NVIDIA drivers are out-of-tree kernel modules and will never ship on the image. The only GPU route under consideration is a Vulkan build of llama.cpp on Mesa's open drivers. It is not built yet, and it would help only discrete cards from the Radeon RX 400 or GeForce GTX 10 generation onwards.

The engine itself is a single package, ai2-llama-cpp, taken from one pinned llama.cpp release. It carries a CPU backend module for each instruction-set level, and at start-up ggml scores them against your processor and loads the best, with plain x64 as the floor. Before release, every file that must run on a pure-SSE2 machine is disassembled and checked, because one stray SSE4.1 instruction would crash an old computer.

For anyone who likes to peek under the lid, the README describes three cooperating parts. The Adaptation Engine detects the hardware, assigns the tier and applies the matching configuration. The Workflow Engine describes what you want to do as declarative YAML profiles, and the Runtime Engine executes models, with llama.cpp as the local runtime. Workflows request capabilities, tiers grant a subset and runtimes execute what was granted. Tier definitions, the model catalogue and workflow profiles are plain data files, and the tool is deliberately small: Python 3.11 or later and PyYAML only. ai-2 workflow install downloads a profile's models and prints its packages as a pacman line rather than installing them.

Models come from a curated catalogue of seven, all Q4_K_M GGUF files that download anonymously, with no Hugging Face account and no licence-gate click: Gemma 3 270M, Qwen2.5 0.5B, Qwen3.5 0.8B, Gemma 3 1B, Qwen3 1.7B, SmolLM3 3B and Qwen2.5 7B. The Qwen and SmolLM3 models are Apache 2.0, while Gemma carries Google's own terms of use. Downloads resume if interrupted and are verified against SHA-256 checksums. The recommendation logic first keeps the models that fit your RAM with headroom, then picks the largest one whose estimated speed is still usable. The ISO bundles Gemma 3 270M, so chat works the moment installation ends, even offline. The project is frank that sub-1B models are fluent but unreliable on facts and arithmetic. One measured curiosity: Qwen3.5 0.8B carries a multi-token-prediction head, but speculative decoding halves its speed on a compute-bound old CPU, so AI-2 leaves it off.

Which brings us to the star of the show: Knowledge Packs. A pack is a set of documents your computer searches and answers from in seconds, offline, naming the source of every answer. No AI writes those answers. You see the documents' own passages, so an old machine gives exactly the same results as a new one, and nothing can be made up. On the machines AI-2 is built for, that beats a chat reply, which can take minutes. Three packs ship on the image: AI-2 Help, Linux Essentials with twenty everyday tasks, and Everyday Reference covering 196 countries with their capitals, currencies, languages and calling codes. Ask through Applications, then AI-2, then Search Knowledge, or with ai-2 doc search. Type a result's number and the full document opens in the w3m text browser at the right passage.

You can make your own, too. Index your PDFs and notes on a fast computer with ai-2 doc index, export them as one pack file with ai-2 knowledge export, and install it on the old machine, which then searches them offline.

Searching your own documents relies on an embedding model, a different animal from the chat models and never offered for chat. AI-2 picked Nomic Embed Text v2 MoE, multilingual and 345 MB, for Standard and above after it found the right part of a Spanish test document for all twelve questions. The 85 MB English Nomic Embed Text v1.5 serves Tiny and Light and managed eleven of twelve. The earlier favourite, all-MiniLM-L6-v2, found only four and was dropped. The cost is speed: on the project's 2016 reference laptop either model embeds about nine tokens a second, so indexing there is a start-it-and-come-back job.

A community catalogue on GitHub, called ai2-knowledge, lists every pack with its download, and anyone can share one by pull request, after automated checks download the file, install it and compare it with what its catalogue entry claims. There is even ai-2 gopher, which lets one computer serve its packs to every other machine on the network, readable by any Gopher client with no AI-2 needed on the asking side.

For conversation, ai-2 chat starts the local AI and opens a chat page in your browser. The address bar will say "Not Secure", which is expected: the page comes from your own computer at 127.0.0.1, so nothing leaves the machine. The AI stops by itself when idle, or on demand with ai-2 stop. 

On low-end hardware, ai-2 chat --terminal is the faster choice, since it skips the browser and works over SSH. 

Programs can use ai-2 serve, an OpenAI-compatible endpoint on port 8080, with a companion embeddings endpoint on port 8081. Its documentation lists Open WebUI, Paperless-GPT, Open Notebook and Blinko as compatible according to their own upstream documentation, though it says plainly these have not yet been driven end to end from an AI-2 box. It is a foreground command with no boot-time service yet. Weaker machines can act as clients, running ai-2 chat --remote against a stronger one. Speech lovers get ai-2 transcribe, which turns a recording into text locally using whisper.cpp.

Trust is built in. Every AI-2 package is signed with the project's key and pacman is set to require signatures, so the signed repository is checked on every update. A wiki page called Verifying and Pinning explains the key, how pacman checks packages and how to stay on an older version if a new one disagrees with you. Model downloads are checked against catalogue checksums and keep to HTTPS through redirects, and since 0.19.0 ai-2 remote set no longer sends an API key over plain HTTP.

Accessibility receives candid treatment too. Terminal chat prints whole sentences, one per line, which screen readers follow naturally, and --speak reads each answer aloud. A single ai-2 accessibility setup installs Orca and the speech stack. The wiki is equally upfront that installation still needs sighted help, because the graphical installer runs as root and is invisible to Orca.

Getting Started

Download the ISO from the GitHub releases page; it is mirrored on SourceForge. Verify it with sha256sum -c ai-2-x86_64.iso.sha256. The checksum is itself signed with the same key that signs every AI-2 package, key F1889E37B4E5FEC8, so you can check that with gpg too. Write the image to a USB stick of 4 GB or more, using dd on Linux or macOS, or Rufus or balenaEtcher in default mode on Windows. Mind the device name: dd erases whatever it is pointed at.

Boot from the stick. A live desktop logs in by itself, so you can try AI-2 without touching your disks, and a START-HERE guide explains the installation. The boot menu also offers "Check this stick first" to verify the written stick, and a safe-graphics entry for screens that stay black. Remember Secure Boot must be off. Choose "Install AI-2" when ready, and the setup wizard greets you at first login. 

Anyone can add a language to the installer through the TRANSLATING guide, with nothing to compile.

Already running Artix or an Arch-based system? Add the signed [ai2] repository, import the signing key, then install ai2-keyring, ai-2 and ai2-llama-cpp and run ai-2 wizard.

The commands you will use most are friendly. ai-2 detect shows what AI-2 sees, ai-2 tier explains your tier, ai-2 recommend suggests a model, ai-2 model pull fetches it and ai-2 chat starts talking. 

ai-2 doctor checks engine, model, tuning, services and repository key, while ai-2 report writes a file to attach to bug reports. 

In total there are 27 commands, all documented in the wiki. Updates arrive through ai-2 update or plain sudo pacman -Syu, covering AI-2, the engine, the model catalogue and the system together. 

Because AI-2 is lean by design, shipping only a desktop, browser, text editor and the AI engine, anything else is a pacman -S away.

Stuck? In the live session, double-click "Remote help (SSH)" and it shows the exact command and password a helper types on their own computer. Only machines on your own network can connect, and the access ends at shutdown. On an installed system SSH is off by default, and the FAQ shows the one command that switches it on.

Questions belong in GitHub Discussions, and confirmed bugs in Issues. Documentation lives in the wiki, and testers on old hardware are warmly invited to report their scores.

A closing word. AI-2 will not write your novel in a blink on a 2011 laptop, and it never pretends otherwise. Its charm lies in measuring first, promising second and keeping the useful parts, from searchable Knowledge Packs to a private chat, working on hardware others have written off. It is early days, but it is an honest and thoughtful start.

Disclaimer: AI-2, Artix Linux, Arch Linux, llama.cpp, Qwen, Gemma, SmolLM3, Xfce, Windows and all other names mentioned are trademarks or trade names of their respective owners, and no affiliation is implied. We strive for accuracy using official sources, yet projects evolve quickly, so please verify details with the official channels. Use open-source software responsibly and legally, and respect each project's licence.

References

• AI-2 on GitHub

• AI-2 Wiki

• AI-2 Releases

• AI-2 on SourceForge

• AI-2 at ProWoos

• AI-2 on DistroWatch

---

🤖🧠🇦🇮👾

↑ Back to top