Self-initiated demo by Flow Lab. Candlewren and Lantern One are invented: nothing here is for sale.

Candlewren

Lantern One

An AI computer that keeps your work in the room.

Lantern One runs large language models on your desk. Contracts, patient notes and unreleased work are read, summarised and searched without leaving the building.

Reserve a unit See inside

Four layers, one quiet box

  1. Glass shell

    Low-iron glass over a recycled aluminium frame. The light of the core shows how busy the machine is, so you never need a status screen.

  2. The core

    A neural processor with 128 GB of memory shared by every model you load. Brighter light means more work in progress.

  3. Compute board

    Storage for your documents and models, and two network ports that you can switch off for fully offline work.

  4. Silent cooling

    A copper vapour chamber and one slow fan keep it at 31 dB under full load, about as loud as a quiet library.

Built for long days of work

Unified memory for models and documents
128 GB
Parameters in the largest model it runs locally
70 B
Noise at full load
31 dB
Peak power, from a normal wall socket
240 W

Nothing leaves the room

Cloud AI means sending your files to someone else's servers. Lantern One reads them where they already are.

  • Works offline

    Unplug the network cable and everything still works: chat, search, summaries and transcription.

  • Your models stay yours

    Fine-tune on your own documents. The weights are stored on the device and can be exported or wiped at any time.

  • A log you can read

    Every question, file and answer is listed in plain language, so an auditor can see exactly what the machine did.

Made for work you cannot upload

  • Law firms

    Search ten years of case files and draft first versions without privileged documents reaching a third party.

  • Clinics

    Turn consultation recordings into notes on a machine that stays behind the clinic's own door.

  • Studios and publishers

    Summarise manuscripts and scripts before release, with no risk of them turning up in someone's training data.

Reserve a Lantern One

First units ship in spring. Reserving costs nothing, and you can cancel with one email.

$4,800 per unit, with three years of updates

Questions

Which models can it run?

Open-weight language models up to about 70 billion parameters, speech-to-text and image models. New models install from a USB drive or, if you allow it, over the network.

Does it need an internet connection?

Only for software updates, and you can install those from a USB drive instead. Everyday work happens entirely offline.

How many people can use it at once?

Around twenty people chatting at the same time, or five running long document jobs. Teams bigger than that usually add a second unit.

What happens to the data if we return it?

You wipe it from the settings screen before sending it back, or keep the storage module and we replace it at no cost.