A private AI appliance is a dedicated physical device that runs an AI assistant
entirely on its own hardware. The language model, your conversations, your memories
and your documents all live inside the box, so it works with the internet unplugged,
needs no account, and has no subscription. It is the hardware alternative to cloud
chatbots like ChatGPT and to cloud smart speakers like Alexa — you trade some raw
intelligence and speed for ownership and privacy you can verify yourself.
PHNTM One is one example: a $549 desk device built on a Raspberry Pi 5
that runs Google's open-weights Gemma 3 4B model on-device. This guide explains the
category, compares it honestly with the alternatives — including the free
do-it-yourself route — and lists what to check before you buy any device like this,
ours included.
THE SHORT ANSWER
what it is
A standalone computer with a local AI assistant built in — an appliance on your desk, not an app or a cloud service
what makes it private
The model runs on the device. Nothing you say has to leave your house to get an answer
who it's for
People who want an AI assistant for notes, memory, reminders and documents, and won't send that to someone else's servers — and who don't want to build and maintain a local LLM setup themselves
the trade
A small on-device model is less capable and slower than a frontier cloud model. That is physics, not marketing
typical cost
One-time hardware price, no monthly fee. PHNTM One is $549 once
The four ways to get an AI assistant, compared
If you want an assistant that answers questions, remembers things and helps you
through the day, there are really four options. Each one is the right answer for
somebody.
1 · CLOUD CHATBOT APPS — ChatGPT, Claude, Gemini
how it works
Every message goes to the provider's data center and the answer comes back
strengths
By far the smartest and fastest option. Free tiers exist
costs
Typically $20+/month for the good models. Needs internet and an account. Your conversations live on their servers under their policy
best if
You want maximum intelligence and are comfortable with your data leaving your house
2 · CLOUD SMART SPEAKERS — Alexa, Google Home
how it works
An always-listening microphone waits for a wake word, then sends audio to the cloud
strengths
Cheap hardware, instant answers, controls lights and other smart-home gear
costs
Stops working without internet or if the service is shut down. The microphone listens for the wake word whenever it isn't muted
best if
You mainly want smart-home control and timers, and the privacy trade doesn't bother you
3 · DIY LOCAL AI — Ollama, LM Studio, GPT4All, Jan on your own computer
how it works
Free software that downloads an open-weights model and runs it on a PC or Mac you already own
strengths
Free, private, and a good GPU or a recent Mac will run bigger, faster models than any small appliance
costs
It's a chat window, not an assistant: long-term memory, reminders, voice, document search and phone access are projects you assemble and maintain yourself. It ties up the computer it runs on
best if
You're technical, you enjoy the tinkering, and you already have strong hardware. Genuinely: if that's you, do this
4 · A PRIVATE AI APPLIANCE — a dedicated device like PHNTM One
how it works
A purpose-built box with the model, memory, voice and screen already integrated. Plug it in and use it
strengths
Private like DIY, finished like a product. Always on, on its own hardware, with its own screen and voice. Works offline. No account, no subscription
costs
Up-front hardware price. A small on-device model is slower and less capable than the cloud
best if
You want the privacy of local AI without building it, and you value a companion that remembers over a genius that forgets
There is a longer, more pointed version of this comparison on the
compare page — including the rows where the cloud wins.
What to check before you buy any private AI device
"Private" and "local" are easy words to print on a box. These are the questions
that separate a device that is private from one that merely says so. Ask them of
any product in this category — including ours.
- Does it work with the internet unplugged? Not "most features" — the actual assistant. If answers stop when the router does, the intelligence isn't in the box.
- Can you verify what leaves the device? A privacy policy is a promise. A live connection log on the device, checkable with standard tools like
tcpdump, is evidence.
- Does it need an account? An account means a server, and a server means your device depends on the company staying alive and staying honest.
- When is the microphone live? Always-listening wake words and push-to-talk are very different privacy postures.
- What happens if the company shuts down? A device that depends on the maker's servers becomes e-waste. A device that doesn't, keeps working.
- Can you get your data out — and delete it for real? Look for a full export, and deletion that removes the searchable copy too, not just the visible record.
- Can you get in? Documented SSH or equivalent access means the machine is actually yours.
- Are the speed numbers honest? Small hardware is slow compared to a data center. A maker who publishes measured numbers, including the bad ones, is a maker you can believe about the rest.
- What is the total cost? One-time price versus subscription, and whether any "optional" cloud feature is quietly required.
How PHNTM One answers those questions
PHNTM ONE · THE FACTS
hardware
Raspberry Pi 5, 8 GB RAM, 10.1″ 1920×1200 touchscreen, speaker, active cooling, 256 GB storage, enclosed case (
specs)
model
Gemma 3 4B (Q4_K_M) running on-device, plus a local embedding model for memory and document search
offline
Yes — chat, memory, voice, documents and reminders all work with the internet unplugged
what it does
Conversation with long-term memory, one-time and recurring reminders, follow-through on open loops, question-answering over your own documents with citations, voice in and out, and a private chat on your phone over your own network (
how it works)
verification
"Watch the Wire" shows live outbound connections on the device; every claim has a command you can run (
proof,
privacy)
account
None. Zero telemetry
microphone
Push-to-talk only. No wake word, never always-on
if we disappear
Nothing changes — there are no servers to turn off (
continuity)
your data
Full "Take Everything" export; deletion removes the record and the searchable copy
access
Debian Linux, SSH documented, never removed
speed, honestly
About 15–20 seconds per answer when warm; about 90–106 seconds for the first answer after an idle spell while the model reloads
cloud option
"Boosted" mode: optional, off by default, uses your own API key, and every answer that left the box is labelled
price
$549 once, no subscription. Made to order, ships within 7–10 business days, free US shipping, 30-day returns (
order ·
policies)
Who should not buy a private AI appliance
If you need the smartest possible model for coding or research, use a frontier
cloud model — a 4-billion-parameter model on a Raspberry Pi is not in that league
and never claims to be. If you want instant answers, the cloud wins. If you want
your assistant to run your lights and order groceries, buy a smart speaker. And if
you already have a strong GPU and like tinkering, running Ollama or LM Studio
yourself is free and will outperform any small appliance on raw speed.
A private AI appliance is for the person who read that paragraph and still wants
the box: someone for whom where the conversation lives matters more than how
clever the answer is.
Common questions
Is there an AI assistant that works completely offline?
Yes. Any assistant that runs its language model on local hardware works offline. You can do it yourself with free software such as Ollama or LM Studio on a capable computer, or buy a dedicated private AI appliance such as PHNTM One, which runs Gemma 3 4B on a Raspberry Pi 5 and keeps chat, memory, voice, documents and reminders working with the internet unplugged.
Is a private AI appliance a replacement for ChatGPT?
For everyday assistant work — remembering things, reminders, drafting, answering questions about your own documents — yes. For frontier-level reasoning, no: a small on-device model is less capable than ChatGPT's cloud models. PHNTM One handles this honestly with an optional Boosted mode that uses your own API key for hard tasks and labels every answer that left the device.
Why buy a device instead of running a local LLM on my own computer?
If you're technical and already own strong hardware, running it yourself is a great option and costs nothing. An appliance is for people who want the result without the project: memory, reminders, voice, document search, a touchscreen and phone access already integrated, always on, on hardware that isn't your work computer.
Can a Raspberry Pi 5 really run an AI assistant?
Yes, with the right expectations. A Raspberry Pi 5 with 8 GB of RAM runs a quantized 4-billion-parameter model such as Gemma 3 4B entirely in memory. On PHNTM One that means answers in roughly 15–20 seconds when the model is warm. The measured tokens-per-second, RAM and thermal numbers are published on the
proof page.
How do I know a "private" AI device isn't secretly sending data out?
Don't take the maker's word for it — check. Unplug the internet and see whether it still answers. Watch its network traffic from your router or with a tool like tcpdump. PHNTM One also shows its own live outbound connections on the device's screen ("Watch the Wire"), and the
privacy page documents exactly what stays local.
Who makes PHNTM One?
PHNTM One is designed, assembled and tested by Jacob DeCamp, a solo builder in the United States, and sold only at phntmcore.com. It is not affiliated with other companies or products that use the name "PHNTM". More on the
builder page.