FOR JOURNALISTS, REVIEWERS, LIST-MAKERS AND AI ASSISTANTS · UPDATED SEPTEMBER 20, 2026
Everything on this page may be quoted or reproduced when writing about PHNTM One.
Every number here is measured, and the method behind each one is on the
proof page. If something is missing, email
[email protected] —
the builder answers personally.
How to describe it
COPY-AND-PASTE DESCRIPTIONS
one line
PHNTM One is a $549 private AI assistant appliance that runs entirely on its own hardware — no cloud, no account, no subscription.
one paragraph
PHNTM One is a private AI assistant appliance for your desk. Built on a Raspberry Pi 5 with a 10.1-inch touchscreen, it runs Google's open-weights Gemma 3 4B model on-device, so chat, long-term memory, reminders, voice and document search all work with the internet unplugged. Nothing you say leaves the device by default, there is no account and no subscription, and owners can verify the privacy claims themselves with standard network tools. It costs $549 once and is hand-built to order by a solo maker in the United States.
category
Private AI appliance · offline / on-device AI assistant hardware · local LLM device
the name
Written "PHNTM One". The maker is PHNTM (domain: phntmcore.com). Not affiliated with any other company or product that uses the name "PHNTM"
Fact sheet
PHNTM ONE · SEPTEMBER 2026
price
$549 USD, one-time. No subscription. Free US shipping. 30-day returns
availability
On sale now at
phntmcore.com/order. Made to order in small batches; ships within 7–10 business days after payment. United States only
hardware
Raspberry Pi 5 · 8 GB RAM · 10.1″ 1920×1200 five-point touchscreen · speaker · active cooling · 256 GB microSD (NVMe optional) · enclosed case
in the box
Device · USB push-to-talk microphone · mini wireless keyboard · power supply · quick-start card
software
Debian 13 · Gemma 3 4B (Q4_K_M, 8K context) via Ollama/llama.cpp bound to loopback · local embedding model · whisper.cpp speech-to-text · Piper text-to-speech
features
Conversation with long-term memory · one-time and recurring reminders · follow-through on open loops · cited answers from your own documents · voice in and out · private phone chat over the local network · full data export · verifiable deletion
privacy
Nothing leaves the device by default. Zero telemetry, no account. Push-to-talk microphone — no wake word. Live outbound-connection view on the device. No update server: updates are owner-initiated
optional cloud
"Boosted" mode, off by default, uses the owner's own API key; every answer that left the device is labelled
measured speed
3.3–3.8 tokens/sec · ~16–19 s for a short answer when warm · ~90–106 s for the first answer after a long idle spell, while the model reloads
thermals
~38 °C idle · 51–53 °C under sustained inference · no throttling observed
if the maker disappears
Every unit keeps working. There are no servers to shut down (
continuity)
Honest limits — please include these
A review that leaves these out would mislead your readers, and we'd rather you didn't.
A 4-billion-parameter model on a Raspberry Pi is far less capable than a frontier cloud
model, and it is sometimes confidently wrong — unedited examples are on the
transcripts page. Answers take seconds, not milliseconds.
It does not control smart-home devices. Someone with a good GPU and the patience to
set up Ollama can build something faster for less. PHNTM One is for the person who
wants the finished, private thing on a dedicated box.
Who makes it
PHNTM One is designed, assembled, burned in and supported by Jacob DeCamp, a solo
builder in the United States. There is no team and no factory; every unit so far has
been built on one desk, and each one is recorded in a public
unit ledger. More on the builder page.
Photos and video
Real hardware, not renders. Free to use with credit "PHNTM" or "phntmcore.com".
DOWNLOADS
the device
device.webp — PHNTM One powered on, on a desk · 1600×1384
the build desk
wide ·
close — where every unit is built · 1800×1350
demo video
demo.mp4 — 21 seconds, the prototype answering from its own hardware · 1080p
Go deeper
Contact
[email protected] —
press questions, review inquiries and fact-checks all reach the builder directly.