Back to blog
Article

Linux Day 2025: From Cybersecurity to Extreme LLM Self Hosting

From Linux Day 2024 with Flipper Zero to 2025 with an AI chatbot on Raspberry Pi 5: history, anecdotes and extreme self hosting of open source LLM with Ollama.

Stefano Capezzone October 25, 2025
Raspberry Pi 5 with M.2 HAT and cooler for LLM self hosting

The last Saturday of October is traditionally dedicated in Italy to Linux Day, the celebration of the operating system that now hosts virtually all major enterprise systems.

The origins of this exemplary system, which definitively showed the world the full power of the open source model, are characterized by a mix of myths and legends, more or less urban, as befits a project of such great success.

According to the prevailing media narrative, in 1991, a young student at the University of Helsinki, usually dressed with the inevitable backpack, the iconic look of a typical student of that era, named Linus Torvalds, began developing a new operating system as a hobby, a Unix-like system for Intel 386 PC architectures, the top-of-the-line processor at the time.

After publishing his famous post on August 25, 1991 on the Internet (usually without specifying what “Internet” meant, leaving the impression it was the web), the project was wildly successful, immediately gaining approval from the Internet community. The project grew enormously, eventually steamrolling the proprietary Unix systems of giants like IBM, HP and Sun Microsystems, becoming the unique and standard system that exists today.

The story I remember is somewhat different, partly because I was there at the time and, as a young recent graduate, I was newly employed at a Roman company, very high tech, heavily engaged in European-scale technological innovation projects.

In the Esprit projects environment of the ’90s, Internet connection was a kind of obligation, also because in companies and organizations that were part of that network, it was fashionable to communicate via email rather than fax (does anyone still remember what fax was?).

The Internet as we know it today was born only in 1996, with the advent of ISP providers. Five years earlier, if you wanted to be connected to the Internet, you went in person to the CNR (National Research Council), specifically to GARR, the Research Network Harmonization Group, where you had to convince a zealous researcher that you and your company were doing real research and, above all, you had to negotiate the assignment of a domain, understandably in class C, trying to get the least penalizing subnetting possible.

We, for example, obtained after stressful and lengthy negotiations a domain with subnet mask 255.255.255.240 and managed to connect all 12 PCs we had in the office.

You read that right, at the time every PC on a LAN was directly connected to the Internet, just like that, bare and raw, without a firewall, with your development server absurdly also acting as a router, being the only computer always on and capable of doing dial-in on demand.

Well, in ‘91 there was no web as we know it today and especially no social networks. Researchers like me argued on newsgroups. The newsgroups of that time were like the bar or university cafeteria, which stood to academic and industrial research as the sports bar stands to football matches. They were basically the social networks of the time.

Unix was the preferred operating system for researchers and independent, somewhat hippie developers of that era. The industrial alternatives were monsters like IBM CICS Cobol DB2, PDP 11, VAX, and shortly after Windows NT server for SMEs.

Unix was an entire ecosystem, based on the open source project from UC Berkeley, but already declined into a myriad of dialects and distributions burdened by intellectual property rights.

In the research sector, we virtually quarreled over some new kernel projects, the heart of an operating system, to bring Unix to Intel PC architectures.

The most important dispute centered around a project called Minix based on the principle that a kernel should be as small as possible (mini-kernel) and everything else should be implemented as an additional and potentially optional module.

And then in August 1991, right on the comp.os.minix newsgroup, appeared a post from this somewhat arrogant Nordic researcher, I remember he was already a researcher not a student, announcing his “small operating system project”.

It wasn’t a tiptoed announcement, steeped in modesty, as the title might suggest. It was a real provocation. On the official and sacred newsgroup of the Minix project, a true heresy was published: a monolithic and quite chunky kernel.

The discussion remained heated for many years, but then reality settled the matter. The faction that wanted tiny kernels, to which I belonged with conviction, had to capitulate before the fact that making operating systems with a monolithic kernel is much easier.

With the undoubted value of the Unix model, given the strong need for a server operating system based on the ever-expanding Intel processor family, and considering the power of the open source model, Linux, with its obese but functional kernel, won, becoming the very emblem of open source.

Actually, the Minix model is the one that truly won in the long run. Today’s Linux has a modular kernel (fancy that) and the only proprietary alternative to Microsoft Windows operating systems is MacOS, which is based precisely on the original Minix project (small world).

The Linux Day Invitation

All this long initial digression to explain that when, unexpectedly, some years ago I received an invitation from Engineering students at Tor Vergata University in Rome to propose a talk at the Rome Linux Day they organized, I could only accept enthusiastically.

Every year Linux Day has a reference theme. That time the theme was the Internet of Things.

Fastal had participated in several IoT projects and was a partner of the initiative born at the Digital Innovation Hub in Rome to create the local The Things Network community.

Foolishly, forgetful of my past as a student and researcher, I showed up with a talk about a real project, with a boringly corporate angle.

The result was that the two classrooms next to where I was performing as a speaker were full, while my audience consisted of about thirty participants, relatively older.

Despite considering this first experience of mine a real flop, I was again invited as a speaker for the following edition.

Linux Day 2024: Hacking with Flipper Zero

The theme this time was cybersecurity. I decided to make up for the previous year’s flop and take a sort of revenge. I thought long and hard, no longer as an entrepreneur, but as a young nerdy student.

At Linux Day 2024, I showed up armed with a small and fun device: the Flipper Zero, complete with Wi-Fi card. The title of my talk was a marketing masterpiece: “Wireless Network Hacking with Flipper Zero and Kali Linux”.

Kali Linux was actually thrown in to better contextualize the talk with the day’s theme. It’s a serious product we actually use for penetration testing and not something easy to show in 20 minutes, but the Flipper Zero is a not very dangerous but very fun and educational toy.

When I entered, the room was already full. During the 20-minute show, the public Wi-Fi network of Room 1 had become a real carnival. Even the ethereal space where we see available network names on our devices had been rendered unusable. Instead of network SSIDs, the words of a famous song scrolled by, then the classroom network name appeared a thousand times, but connecting led to fake pages that looked like those of major streaming providers or e-commerce sites. Laughter and total success.

I believe that since then, no one present in that room has ever connected their device to an open or public network again.

Linux Day 2025: Extreme LLM Self Hosting

Therefore, once invited to Linux Day 2025, whose theme sounded less exciting: “self hosting”, what could I bring?

After a long selection of discarded ideas, an original one emerged: extreme LLM self hosting.

In the spring of 2025, open source (or open weight) LLM models were beginning to establish themselves.

New optimization, quantization and distillation techniques for LLMs and the use of new System 2 and Mixture of Experts architectures had enabled the release of particularly performant small-sized models.

Open Source LLMs represent a very interesting opportunity for companies typically in our target market, because they cut operating costs, prevent vendor lock-in and above all can be deployed within conventional data centers, with minimal upgrades in consumer-grade GPUs, guaranteeing 100% privacy and GDPR compliance.

Running an open LLM on a node with a 5090 GPU and 32 GB of conventional RAM is easy, but how far can we push the downsizing of hosting?

Here’s the challenge: run an LLM and implement a generalist AI chatbot on a Raspberry Pi.

Obviously, some RAM is needed, so the minimum requirement for the experiment is the new Raspberry Pi 5 with 16 GB of RAM, a little gem for about 100 Euros.

The Live Demo

Said and done. At the Linux Day 2025 edition, I presented the experiment live.

The room was packed; I’m starting to have my own fan base.

On the table, a Raspberry Pi 5, packaged with the official fan cooler and expansion HAT to have a 512 GB SSD as mass storage.

On board, an installation of Ollama and Open WebUI, configured to perfection, with a nice collection of 4-bit quantized models, including the incredible Qwen2.5-vl and Qwen3 from Alibaba.

How to connect to the ChatBot? After the previous year’s show, I could never have connected to the classroom network we had hacked and mocked for being a total sieve.

The solution was a mini on-premise Wi-Fi network configuration, with an AP implemented on a small Samsung SoC device, obviously well protected and configured. A single point of vulnerability, attackable with the only technique that knows no technological protections: social engineering.

The network had an apparently bizarre SSID: Wintermute_Mainframe. To connect, a password that I pretended to keep secret (otherwise 200 of you would connect), but I threw in a cryptic phrase, seemingly by chance, “I hope no one catches the reference”.

The ChatBot really looked like ChatGPT, responded fluently, appropriately, and in a few seconds rattled off the Python code to solve the Tower of Hanoi puzzle with a recursive algorithm.

An AI comparable to ChatGPT from 2023 was running without requiring a dedicated nuclear reactor, on a device connected via a USB-C charger to a regular electrical outlet, just slightly warm.

After a few minutes, someone caught my hook and understood. Cyberpunk culture is still alive. They got into the Raspberry Pi and issued the shutdown command. There are promising young people in our university classrooms.

Linux LLM Raspberry Pi Ollama Open Source Self Hosting

Enjoyed this article?

Contact us to discover how we can help you with our solutions.

Contact us