28 Aug 2026
ReachedMVP complete
The first complete LionBox experience enters testing and improvement.
01LionBox roadmap
Building in public. Shipping with purpose.
Follow the path from an idea to the first LionBox production batch. Dates reflect our current plan and will be updated as the project moves forward.
7 Nov 2026
Marketing campaigns begin
We start introducing LionBox to teams looking for a practical private AI alternative.
027-13 Dec 2026
Founders Edition launch week
Orders open on lionbox.ai for Founders Pass holders. Only 50 Founders Edition units will be made.
03Jan 2027
First deliveries
The first batch is expected to ship from late January 2027, in reservation order.
04Jan 2027
Second batch opens
Orders open for the second production batch, following the Founders Edition.
05Road to launch
Private AI, getting closer.
We are building and testing LionBox for its first 50 owners, ahead of launch week on 7-13 Dec 2026.
71%
Estimated build progress · Updated 2 November 2026
What we are building
- 01
Private AI for your whole team
Chat, documents and answers on your own network, with nothing leaving the building.
- 02
Set up and run from your phone
Connect your LionBox with one scan, then add people, share documents and check its health from the app.
- 03
Looks after itself
Safe updates, automatic recovery and a box that keeps working without the internet.
Still to come before the first deliveries: final testing, certification and building the first batch.
Estimated from our launch checklist, with work in progress counted as half done.
Read the development logProject updates
Development log
Notes from the build, published newest first. This log will grow through testing, launch and production.
See how busy your LionBox is
The app now shows how many answers the box is working on and how much conversation memory is in use, so you can tell at a glance when your team is pushing it hard.
From chat to finished files
A first version of Create file turns a chat into a presentation, a document or a spreadsheet that opens in Office, built entirely on the box. We are now polishing layout and design.
Better answers, longer conversations
We built a benchmark of 65 real business tasks and used it to tune how LionBox reads documents and handles long chats. A newer engine in testing remembers about 1.6 times more conversation and answers faster when many people chat at once.
Company, team and personal documents
Share documents with the whole company, with one team, or keep them private. Admins and team managers decide who can edit what, and every document stays on the box.
Set up with a single scan
Owners no longer type pairing codes. After signing in, your phone connects to your LionBox automatically, with the QR code on the box as the only fallback.
Built to keep running
LionBox now recovers on its own when part of the system stops, shows a friendly maintenance page instead of an error and protects itself in a safe mode until everything is healthy again.
A chat that feels like LionBox
The chat your team uses got the LionBox look, light and dark themes, and starter prompts written for everyday office work instead of generic examples.
The new lionbox.ai
The full website went live in preview, with product details, how it works, pricing and the Founders Pass reservation for the first production batch.
Load-tested for a whole office
We sent about 2,000 messages through one box from simulated colleagues, including everyone pressing send at the same moment. No request failed, the first words typically appeared in 0.07 seconds, and the box had headroom for a 20-30 person team. The test also caught a crash when several people pasted long documents at once, which we fixed and tested again.
Orders and invoices
LionBox Cloud learned to take orders, check EU VAT numbers and send invoices, ready for the first customers.
Think before answering
A Think switch lets the 9B model work through a problem step by step before it replies. On our reasoning test set it raised correct answers from 83% to 95%. It takes longer, so it stays off until a question needs it.
The MVP is ready
The core hardware and software experience came together, moving LionBox into structured testing and improvement.
Speech to text, on the box
Voice notes and recordings are transcribed on the box's processor rather than its graphics card, so dictation never slows anyone's chat. A ten-minute recording is ready in about 16 seconds, roughly 37 times faster than real time.
LionBox Cloud and support
The cloud side came together: accounts, licences, a registry of every box, signed update delivery and a support desk for help and warranty cases.
A box that looks after itself
The software that runs the box gained health monitoring, safe updates that roll back if something goes wrong, network changes that undo themselves if they cut you off, and a power schedule.
Bluetooth setup and rescue
We designed the Bluetooth link between the phone and the box. It handles first setup and can repair the box's network settings when the office network is down, while models and updates always travel over the much faster local network.
Reading whole documents
Most chat tools cut an upload into a few snippets and hope the right ones get picked. LionBox reads a document in full whenever it fits and only searches inside very large files. In our tests, correct answers about uploaded documents rose from under half to about two thirds.
Designed for people who aren't IT
We shaped the owner app around one rule: anyone in the office should be able to set up and look after LionBox without technical words, logs or service lists.
Our own build of the 9B model
Rather than run the public release as it comes, we prepare the 9B model ourselves, tuned layer by layer for the LionBox graphics card. Compared with the public build, ours gives almost twice as much room for conversations and documents and answers about 20% faster, with no measurable loss in quality across 800 reviewed answers.
A test bench for every change
We built an automated test bench that every model, setting and software change has to pass before it reaches a box: general knowledge, maths, several languages, tool use, reading images and finding one fact buried deep in a long document. Results are compared question by question, so a regression can't hide in an average.
Private by design
We set the rules everything else follows: LionBox works without the internet, conversations and documents never leave the box, and the cloud only handles accounts, licences and updates.
One graphics card, a whole team
We tuned the AI engine to serve many people from a single graphics card: requests from different colleagues are processed together instead of one after another. Up to 16 people can press send at the same moment and every one of them sees the first words in under half a second.
Choosing the graphics card
We compared NVIDIA and AMD graphics cards on speed, memory and software maturity. A 16 GB NVIDIA card won: the AI software we rely on runs fastest and most reliably on it, while the alternative still had open stability problems.
Choosing the 9B model
We tested open models from 8 to 14 billion parameters on one graphics card for answer quality, speed, languages, tool use and reading images. A 9B model gave the best balance: strong answers in many languages, reliable tool use and image reading, with room left on the card for long conversations. Larger models don't fit on a single card with room to spare.
The first prototype
We built the first prototype from desktop parts: one graphics card, Ubuntu Server and an open-source AI engine, running with no screen attached. It became the reference box that every later benchmark, load test and update runs on.
Product development began
Work started across both sides of the product: the appliance hardware and the software experience around it.
First measurements
Before designing anything, we measured what one desktop graphics card can do with today's small open models. It wrote answers about eight times faster than anyone can read, for several people at once, which made a quiet office AI box worth building.
The idea for LionBox
The initial idea arrived and research began into how private, local AI could become practical for small teams.
Follow the build