We build small, specialised models you host yourself - in a browser, on a laptop, on-prem, in your private cloud, inside your app or on the edge. Your data never leaves. The weights are yours.
In development · public site coming to ai.lamplitlabs.com
Sovereign means you hold the weights, you run them on hardware you control, and nobody else decides when the model changes, what it costs or who reads the traffic.
Data stays homePrompts, documents and answers never cross your perimeter. There is no third party to audit because there is no third party.
You hold the weightsThe model is a file you own, under the licence of its open base. Copy it, archive it, run it in ten years.
It changes when you say soNo silent model updates and no behaviour drift between Monday and Tuesday. A new version is a decision you take, measured against the old one.
No meter, no outageOnce it is on your machine there is no per-token bill, no rate limit and no status page to watch. It works offline.
Where it runs
One model, every place you work
In the browserNothing installed, nothing uploaded - the model downloads once and stays in the browser cache.
On a laptop or desktopA native runtime on the machine that already has the data. Works on a train, on a plane, behind a firewall.
On-prem or private cloudAn inference endpoint inside your network, behind your identity provider, on hardware you already own or rent.
Inside your appShipped with your product as a file. Your users get the feature without an account at a third party.
On the edge and IoTSmall enough for gateways, kiosks and industrial controllers where the round trip to a cloud is the problem.
Models
What you can take home today
Each model does one job. Each one ships only after it passes every release gate on a 177-scenario suite, and these are the numbers it was measured on.
German tutor · browser edition
Light German Tutor v17
A German tutor for A2-C1 learners and the telc / Goethe exams that runs entirely in your browser: the weights download once, nothing you type leaves your machine.
Parameters
3 billion
Facts
92.1% (gate 85.0%)
Runs in
Browser (WebGPU in Chrome or Edge) · desktop · server
Size
2.1 GB download (q4) · 3.6 GB (q8)
Licence
Apache-2.0 (inherited from the base model)
8/8 release gates passed · measured 12 September 2026
The same tutor at 8 billion parameters for a GPU server or a workstation: fuller explanations, C1 transformations and exam preparation, measured on the same suite.
Parameters
8 billion
Facts
92.7% (gate 85.0%)
Runs in
GPU server · workstation (PyTorch)
Size
15.6 GB of weights · about 20 GB of memory
Licence
Apache-2.0 (inherited from the base model)
7/7 release gates passed · measured 15 September 2026
Numbers from the Lamplit Light release gates as of 16 September 2026. Each model's date is when its gates ran; a newer version replaces it here once it passes.
Have a job for a model?
Tell us the task, where it has to run and what a right answer looks like. We will tell you whether a small model can do it - and show you the measurement before you sign anything.