Air-gapped by default
No account, no cloud key, no telemetry. Hearth ships with its radios ready but idle — the moment you want, pull the physical disconnect and it answers from memory alone.
Hearth One · now shipping
Home is the
firewall.
Hearth is a quiet AI box that runs open models on your own network. Ask anything, share files, talk to it — and every byte stays inside your walls.
- No cloud, no accounts
- 18 W, whisper quiet
- Works offline
Why Hearth
Everything an assistant knows,
stays at home.
Hearth runs open-weight language, vision and speech models on hardware built for the living room — not the data center.
Knows your house
Point it at recipes, invoices, baby monitors or the garage manual. Hearth indexes your files on-device and answers with your context — not a stranger's internet.
Quiet, in every sense
A fanless 18 W design sips less power than a router and hums less than a library. Leave it on the shelf; you'll hear the kettle first.
Every screen, one brain
Phones, laptops and the TV talk to Hearth over your Wi-Fi. One private brain, every room — with per-room memory you can wipe in a tap.
Open models, yours to swap
Choose from a curated shelf of open-weight models, or sideload your own fine-tune. No subscription gates the best one.
Fail-closed privacy
If a request ever tries to leave, a hardware watchdog blocks it and lights the ember ring red. You always know when something is outbound. (It won't be.)
Hearth · living room · offline
no uplink
model: hearth-8b · 42 tok/s
0.0 s
A real exchange, replayed. The only network Hearth ever uses is yours.
1
Plug it in
One cable, one minute. Hearth joins your Wi-Fi and introduces itself to your screens.
2
Point it at your life
Drop in folders, photos or the drawer of manuals. Indexing happens on the device, overnight.
3
Just ask
From any room, on any device. Answers come from your home, not somebody's server farm.
Technical specifications
Small box. Serious silicon.
Hearth One in ceramic white or basalt black. Numbers below are for the base configuration.
- Compute
- Hearth H1 NPU · 38 TOPS
- Memory
- 16 GB LPDDR5X, unified
- Storage
- 512 GB NVMe, expandable via the back bay
- Models
- Up to 34 B parameters quantized; 8 B at full speed
- Speech
- Far-field mic array, on-device wake word, 6 m range
- Connectivity
- Wi-Fi 6E, 2.5 GbE, USB-C ×2, HDMI 2.1
- Power
- 18 W idle · 65 W peak · silent fanless shell
- Size & weight
- 18 × 18 × 9 cm · 1.4 kg · ceramic-coated aluminum
- Privacy hardware
- Physical network cut-off switch, tamper-evident seal
- In the box
- Hearth One, braided USB-C cable, quick-start card
It answered a question about our heat pump using the manual that's been in a drawer since 2019. Then it filed itself under Utilities. I checked the router logs — nothing left the house.
We turned ours off for a week to see if we'd miss the big assistants. We didn't. The kids ask it for homework help and I don't have to think about where those questions end up.
I fine-tuned a model on my own research notes in an evening. Try that with a subscription.
Pricing
Buy it once. It's yours.
No subscriptions for core features. Ever. Optional HearthCare covers accidental damage and priority support.
Ember
The starter flame
$249
one-time
- Hearth One · 8 GB / 256 GB
- 8 B chat & voice models
- Whole-home assistant
- 2-year warranty
Most loved
Hearth One
The full hearth
$449
one-time
- Hearth One · 16 GB / 512 GB
- Up to 34 B reasoning models
- Vision: documents & rooms
- 4-year warranty
Studio
For tinkerers & teams
$899
one-time
- Hearth Pro · 64 GB / 2 TB
- Fine-tune & swap models freely
- Rack-mount & dual PSU option
- 4-year warranty + priority help
Ships spring 2027 · $20 donated to open-model projects with every preorder · Prices include VAT where applicable.
| Compare | Ember $249 |
Hearth One $449 |
Studio $899 |
|---|---|---|---|
| Memory / storage | 8 GB · 256 GB | 16 GB · 512 GB | 64 GB · 2 TB |
| Largest model | 8 B | 34 B | 70 B+ |
| Voice & wake word | Yes | Yes | Yes |
| Document vision | — | Yes | Yes |
| Sideload your own models | — | — | Yes |
| Warranty | 2 years | 4 years | 4 years |
FAQ
Asked, answered, kept at home.
No. Inference, file indexing and speech recognition all run on the device. Hearth has no cloud service to talk to — the only network traffic is your own devices reaching it over your LAN. If you enable an optional weather or news skill, Hearth shows you exactly what will be sent before it sends anything, and the ember ring glows while it does.
Hearth One runs open models up to 34 billion parameters, quantized to fit its 16 GB of unified memory. For everyday questions, writing, summarizing your documents and controlling your home, most people can't tell the difference. For frontier-level reasoning you'd still reach for a workstation — or Hearth Studio with 64 GB.
Nothing changes. Hearth never needed the internet in the first place. Chat, voice, document search and home control keep working from local memory. Skills that genuinely require the outside world — like package tracking — pause politely and pick up when you're back online.
Yes. HearthOS ships with a curated shelf of open-weight models you can install in two taps, and Studio unlocks full sideloading: drop in any GGUF or safetensors file that fits in memory, set it as default, or route specific rooms to specific models. Your fine-tunes stay on your device.
The aluminum shell is fanless and passively cooled; under normal chat loads it draws about 18 W — roughly a third of a typical router setup — and peaks at 65 W during long reasoning runs. There are no moving parts, so there is nothing to hum, click or spin. Measured noise: 0 dB.
Preorders from this page ship in order starting spring 2027. Your card is charged when the box leaves our workshop, not before. Until it ships you can cancel with one email; after delivery you have 60 days to send it back for a full refund, no questions asked.