Private by design
Inference happens on the silicon in your house. Hearth has no cloud account to sign up for and no telemetry to switch off — because there is none to switch on.
Introducing Hearth
AI that lives
at home.
Hearth is a quiet, beautifully built box that runs powerful language models entirely on your own network. Ask questions, draft, summarize, automate — your words, your files, your home. Nothing ever leaves the house.
- No cloud account
- Fanless, <18 dB
- Ships March 2027
We kept asking where our conversations were going. Hearth is the answer we wanted to hear: nowhere.
Maya Brandt · Co-founder & Chief Engineer, Hearth Labs
Why Hearth
Serious intelligence.
Zero exodus.
Everything a home assistant should be — fast, capable, always on — and nothing it shouldn't be. No accounts, no telemetry, no uploads.
Whisper quiet
A fanless aluminum monolith with a vapor chamber. Under full load it measures below a library's ambient hum — 18 dB(A) idle, and it never spins up a fan, because there isn't one.
Answers in a breath
The Ember NPU streams a 27-billion-parameter model at up to 66 tokens per second. First token lands in well under half a second, so conversations never wait on a data center.
Works when the internet doesn't
Models, documents and automations live on local NVMe. Pull the ethernet cable, drop the Wi-Fi — the kitchen still gets its answers. Updates arrive when you're back online.
One box, whole household
Per-member profiles with their own memory, voiceprints and permissions. Six people, six assistants, one hearth — each one kept strictly to its own business.
Grows with you
The Ember model library adds new open models monthly — chat, vision, transcription, embeddings — verified and signed, installed with one tap, still fully offline after.
Under the hood
Engineered like a
piece of furniture.
Tuned like a server.
Compute & memory
- Neural engine
- Ember NPU, 12 cores, 96 TOPS INT8
- Processor
- 8-core Hearth H1, 3.4 GHz
- Memory
- 16 / 32 / 64 GB unified LPDDR5X
- Storage
- 512 GB – 4 TB NVMe, user-expandable
Model performance
- Largest model
- up to 70 B parameters (4-bit)
- Daily driver
- 27 B class, 48–66 tok/s decode
- First token
- < 0.4 s at 8K context
- Context
- up to 128K tokens
Body & connectivity
- Enclosure
- Fanless anodized aluminum, 132 × 132 × 44 mm
- Noise
- < 18 dB(A) idle · 0 dB(A) under load
- Network
- Wi-Fi 6E · 2.5 GbE · Thread border router
- Power
- 27 W max, USB-C PD, 0.3 W standby
Software
- System
- Hearth OS, signed & sealed, OTA updates
- Interfaces
- Companion app, web UI, REST & MCP APIs
- Voices
- On-device STT & TTS, 14 languages
- Warranty
- 3 years, parts and labor included
Decode speed — median tokens / second
Hearth Pro · code66
Hearth Pro · prose48
Hearth Pro · vision41
Hearth · code53
Hearth Mini · prose22
Measured on Hearth OS 1.2 with Ember 27B at 4-bit quantization. Individual results vary with model and context.
Pricing
Buy it once.
Own it forever.
No subscriptions, no seat fees, no premium tier for your own conversations. Pick the memory your household needs.
Hearth Mini
For a single desk or a first taste of private AI.
$349
- Memory16 GB
- Storage512 GB NVMe
- Max model14 B params
- Profiles2 members
- NetworkWi-Fi 6E
- Warranty3 years
Most households
Hearth
The full experience. Flagship speed for the whole family.
$649
- Memory32 GB
- Storage1 TB NVMe
- Max model27 B params
- Profiles6 members
- NetworkWi-Fi 6E + Thread
- Warranty3 years
Hearth Pro
For power users, small offices and tinkerers.
$1,199
- Memory64 GB
- Storage4 TB NVMe
- Max model70 B params
- ProfilesUnlimited
- Network+ 2.5 GbE, dual 6E
- Warranty3 years
Reservations fully refundable until shipping. Estimated delivery March 2027. Free shipping, 60-day home trial.
FAQ
Good questions.
Honest answers.
No. Every model, voice and automation runs entirely on the device. The network is only used for optional things you turn on yourself: fetching new models, software updates, or reaching your own services on the LAN. Unplug the WAN and Hearth keeps working exactly as before.
The Ember 27B family is the daily driver — chat, vision, coding and translation at 48–66 tokens per second on a 32 GB Hearth. Mini handles up to 14 B models; Pro runs quantized 70 B-class models. The Ember library adds new open models monthly, and any GGUF or safetensors model can be loaded manually.
It has no fan, so it never gets louder than the room it sits in. Under full inference load the surface warms gently and the noise floor stays below 18 dB(A) — quieter than a whisper at one meter. Many owners report they only know it's working by the ember glow.
Yes. Each member gets a profile with its own memory, voiceprint and permissions, recognized on the companion app or the optional Hearth Pod speaker. Conversations and personal data stay separated between profiles and never train a shared model.
Nothing leaves the device, ever. Updates are signed, verified on-device and applied in the background with automatic rollback. Your models, documents and memories are untouched, and you can schedule updates or pause them indefinitely. If you export a support bundle, it's built locally and you read every line before it goes anywhere.
No — Hearth hears you from across the room with its four-array mic and answers through a full-range driver tuned for voices. For bedrooms and gardens, the optional Hearth Pod adds far-field mics and a tactile dial; all Pod audio is still processed on the box itself.
Bring the fire home.
Reserve your Hearth today. Fully refundable, shipping March 2027.
Reserve from $349