I constructed my very own Googlebook with a Raspberry Pi, native LLMs, and previous {hardware}

Say what you are going to about ChromeBooks, however it’s onerous to disclaim the software of light-weight laptops that toughen maximum productiveness duties and combine seamlessly with cloud-based apps – all whilst considerably not up to typical Home windows machines. However as a staunch member of the house lab faction, I’m now not very keen on depending on third-party servers for my on a regular basis duties. With Google doubling down on Gemini-powered products and services for the approaching Googlebook, I’m a little at the fence about clanker equipment that now not most effective drain my pockets, but in addition retailer knowledge on corporate servers.

However the factor is, large-language fashions are strangely nice partners for my productiveness wishes – in the event that they’re paired with my self-hosted arsenal and don’t switch knowledge outdoor my community, this is. So, I had an concept about development my very own model of a Googlebook whilst looking at the legit announcement flow. One bathe concept ended in any other, and shortly, I had a DIY Googlebook that includes an difficult to understand Chromium-based distro, runs on my Raspberry Pi, and harnesses the 26B and 35B MoE fashions hosted on my house lab nodes.

FydeOS is the center piece of this challenge

It’s even higher than ChromeOS Flex

From the entirety we learn about Googlebook up to now, it has a ChromeOS-like UI and a powerful emphasis on integrating with the Android ecosystem (and Gemini, however I’ll take on that one in a little). Since I wish to reflect that on my DIY gadget, I’ll desire a identical distro. Should you’ve been distro-hopping for some time, you’d most certainly level to ChromeOS Flex, which is basically ChromeOS for unsupported gadgets. Sadly, it has an enormous problem – it doesn’t natively toughen Android apps, and that makes it horrible for this challenge.

So, I went with FydeOS as an alternative, which is a fork of Chromium and packs numerous cool options for my makeshift Googlebook. For one, it doesn’t require a Google account by means of default, so I will simply slap a neighborhood person and get in touch with it an afternoon. Because it doesn’t even come with Google’s cloud apps, I will get a de-Googled gadget proper off the bat. It additionally helps Android apps along with a Linux subsystem, making it way more flexible than ol’ ChromeOS Flex.

The most productive section? FydeOS plays extremely neatly at the Raspberry Pi. I’ll most certainly sound insane once I say this, however I’ve had higher efficiency whilst operating emulators and looking at movies in this random distribution than the real Raspberry Pi OS. Putting in it used to be so simple as flashing it on a microSD card, plugging it into the Raspberry Pi, and going throughout the interactive initialization wizard.

After I’d gotten FydeOS up and operating, I started arming it with Darktable, VS Code, Krita, and different productivity-centric Linux apps, which labored extremely neatly. However regardless of my aversion to Google apps, I did set up Play Retailer when configuring the Android subsystem. With a controller plugged into my Raspberry Pi, I had 0 problems enjoying mild Android video games, and the similar applies to the PSP titles I attempted operating on an Android-based emulator. Now, hardcore equipment like DaVinci Unravel are surely an excessive amount of for this miniature gadget. However in relation to standard productiveness equipment and light-weight gaming, I don’t have any proceedings in any respect.

Open WebUI + native LLMs supply all of the AI options I may ask for

Seems, high-parameter fashions run neatly even on decade-old machines

Whilst we’re relating to in depth products and services, it’s time to deal with the elephant within the room: AI duties. I’ve attempted operating a host of LLMs on my Raspberry Pi, however it couldn’t move upper than 4B fashions (and no, my experiment involving 9B fashions operating on two SBCs by means of RPC does not depend). However taking into account that even Google’s Gemini-powered equipment depend at the corporate’s servers, it’s about time to carry my house lab nodes into the fray.

Now, I’m now not going to mention that my LLMs can beat the loads of billions of parameters that Gemini’s gigantic cloud servers can procedure. Alternatively, positive fashions like Qwen3.6-35B-A3B and Gemma-4-26B-A4B are getting eerily on the subject of the reasoning prowess in their cloud opposite numbers. However what makes them absolute game-changers is the Aggregate-of-Professionals structure that they’re constructed round. On typical LLMs, I’d have to dump complete layers into the CPU and gadget reminiscence, which might hugely decelerate their token technology speeds, despite the fact that they might technically run on previous {hardware}.

For those MoE fashions, I will merely toss the professionals and different less-used parts on my RAM, whilst leaving the eye weights at the GPU, permitting out of date {hardware} to run the fashions with out main slowdowns. For reference, I’m speaking a few 10-year-old GTX 1080 (with simply 8GB of VRAM) riding the likes of a 26B fashion (Gemma-4-26B-A4B) at 15 token/s. However apart from that gadget, I’ve additionally were given an RTX 3080 Ti on my gaming PC that runs Qwen3.6-35B-A3B flawlessly the use of the similar tweaks. By way of pairing those LLMs with Open WebUI and tossing a few ComfyUI symbol technology pipelines into the combination, I’ve already were given all of the AI-centric options I may ask for from my makeshift FydeOS gadget.

The CrowView Observe 14 serves because the pc chassis

It does have a unusual look, despite the fact that

Thus far, I’ve were given each the distro and AI options accounted for. Alternatively, my setup is only a Raspberry Pi hooked as much as a observe and KB+M, which is a a long way cry from a compact pc. That’s the place the CrowView Observe 14 comes into the image. It’s a cool-looking clamshell chassis with a integrated observe, keyboard, trackpad, battery, and audio system. Necessarily, all of the equipment of a pc, minus the internals – as a result of the ones are equipped by means of the Raspberry Pi.

I’ll be blunt right here: it has a wacky design, with the Raspberry Pi sticking out from the left facet. I’m additionally now not very keen on the 4-ish hour battery lifestyles, whilst the loss of a webcam is simply as a lot of a buzzkill. However taking into account the quirky nature of this challenge, it meshes neatly with the remainder of the setup.

It will not be a Googlebook killer, however it does have some perks

Now, I’m now not going to fake like this unhinged concept can exchange a Googlebook, in part for the reason that Raspberry Pi isn’t sufficient to function a pc, and in addition as a result of we don’t also have many main points on Google’s new gizmo. However truthfully, I’m beautiful glad with this challenge. FydeOS is an amazing distribution that merits extra popularity from tinkerers, particularly because it’s higher for operating Android apps on Raspberry Pi than LineageOS and different Android-based distros. Likewise, reviving elderly graphics playing cards by means of turning them into LLM-hosting workstations has turn into an obsession of mine these days. Whilst my CrowView Observe 14 isn’t slightly the similar as a devoted pc chassis, this setup is greater than sufficient for LLM-aided on a regular basis duties.

CPU

Arm Cortex-A76 (quad-core, 2.4GHz)

Reminiscence

As much as 8GB LPDDR4X SDRAM

Working Device

Raspberry Pi OS (legit)

Ports

2× USB 3.0, 2× USB 2.0, Ethernet, 2x micro HDMI, 2× 4-lane MIPI transceivers, PCIe Gen 2.0 interface, USB-C, 40-pin GPIO header

GPU

VideoCore VII

Beginning Value

$60