Skip to content
leGMLleGML

What sovereignty means when the weights are yours

On-premise deployment is a legal position as much as a technical one. Notes from three procurement conversations.

Try Le Lion

In brief

Your weights, your premises.

3Procurement conversations
0Data leaves your premises
14 BRuns on hardware you own
On siteDeployment

01 · The position

A legal question before a technical one

When a regulated institution buys a model, the first questions are rarely about accuracy. They are about where the model runs, who can see what, and what happens when the contract ends.

Owning the weights changes the answer to all three. The model runs inside the institution, under controls its regulator already knows.

These notes come from three procurement conversations. The questions were the same each time.

Where things live

One way across: the weights, under licence.

With usThe model learns the sectorOn public and synthetic data. Your data is never touched.
Licence
With youThe model runs in your environmentYour data stays there. Your teams can audit it.

The two sides of a deployment.

What buyers asked

Conversation 1Where do the weights run?On your hardware, inside your perimeter. Nothing is called outside.
Conversation 2Who can audit the model?Your own teams, with the documentation and evaluations we deliver.
Conversation 3What happens when the contract ends?The weights you licensed stay where they are.

Paraphrased and anonymised.

02 · How it works

Four steps, under your control.

  1. IAdaptWe train the model on public and synthetic data.
  2. IIDeliverThe weights move into your environment, under licence.
  3. IIIRun on siteYour data stays where it is.
  4. IVAuditYour teams check it with our evaluations.

The trade-off

Capability and control, at the same time.

Capability →
Small generic model, on siteFrontier model, hosted elsewhereSpecialist model, on site · leGML
Control →

Schematic, not measured.

In our words

“Sovereignty is an engineering property, not a slogan.”
Mohamad AlhajarParis

03 · Why it matters

Capability and control, without the trade

European institutions are told they must choose between a capable model and one they control. At the sizes that matter, we do not think the trade is real.

That is the position behind everything we build.

Le LionTry it →APIAPI reference →

Read next

All notes →

Winning OpenAI's privacy hackathon at Station F

Dispatches

Why a 14B model beats a frontier model on French business law

Benchmarks

Building a benchmark with the people who do the work

Benchmarks

What should your model do?

contact@legml.ai