ADC · November 2026
Open-weight models on your own infrastructure
What if our AI provider is switched off overnight?
We want to be less dependent on the US.
This data can’t leave our network, so it gets no AI at all.
Our AI costs go up with every new use case.
Generic AI doesn’t know our work, and we can’t train it.
Some developers burn thousands of euros a month in coding tokens.
What we hear from banks and insurers
Complex reasoning
Bulk, non-sensitive work
Sensitive data, self-hosted
High-volume, narrow tasks
Fast classification
The trained model is published, so you can run it on your own hardware. Nothing goes back to the lab that made it.
For a single, narrow task, a small model is often just as good as the biggest ones.
Download the model, run it anywhere
Up to 100× smaller and cheaper than frontier models
Most AI solutions don’t need the very best model. More than fifteen open models from 2026 score in this range.
A US export directive took Claude Fable 5 offline worldwide, EU included, for 19 days.
Weeks later, GPT-5.6 launched to government-approved partners only.
Microsoft France, under oath: it “cannot guarantee” French citizens’ data won’t go to US authorities.
An EU region doesn’t change who the operator answers to.
Anthropic cut Windsurf’s direct access to Claude with less than five days’ notice, after OpenAI moved to buy it.
Your access depends on the provider’s business interests.
Data to China, where by law companies must help state intelligence
Data to the US, where authorities can demand it, within limits the EU accepts
Data stays in the EU, but US law can still make the provider hand it over
Data and provider both under EU law
Your own model, on GPUs you rent in the EU
Model, data and hardware are all yours
Closer to you, more sovereign
Many organisations todayAbout 2,000 engineers moved from Claude Code to GLM, an open-weight model Revolut hosts itself.
Around $0.30 per million tokens: a tenth of what it paid Anthropic and OpenAI.
A fine-tuned 2B open model checks content against a pricing policy as accurately as the best proprietary models.
Self-hosted on one GPU: 35 ms per answer, against about 900 ms for frontier APIs.
bunq says it A/B tested open model Kimi K3 in Finn’s support chats for a week, on dedicated GPUs in the EU.
bunq claims quality within 1% of the best proprietary model, at around 60% of the cost per answer.
Company-reported figures. Sources: The Register, Airbnb Engineering, bunq on LinkedIn
We ran open and closed models on support tickets, sorting each one by intent, team and priority.
Fine-tuned: trained further on our own labelled tickets
You can only do this when you have the weights, which is a big reason open weights matter.
Billions of weights, like these sliders
Now it answers the way your team would
Fewer than 1,000 labelled examples from your history, and one GPU to train on
Narrow, frequent tasks, where the same question comes up thousands of times
When the task changes, you have to train the model again
Illustrative shape from published break-even analyses. Where the lines cross depends on your model, hardware and prices.
Plan for utilisation: share GPUs across use cases and fill quiet hours with batch work.
Our results match what public benchmarks and other teams have already found.
Out of the box, the best open model did the job about as well as the best closed model we tried.
Fine-tuned on our own past tickets, a small open model beat every model we tried, the frontier included.
The fine-tuned model ran on one GPU and answered several times faster than the frontier.
Often the strongest first step: an assistant for policies, procedures and client files. It answers with its sources and only shows what each person may see. Nothing leaves your network.
For AML alert triage, most alerts are false alarms. A fine-tuned small model closes the clear ones with a confidence score and sends the rest to an analyst.
For coding, an agent works across all your repositories on an open-weight model inside your own network: no per-token bill, and source code never leaves the building.
For KYC, the model reads ID documents and proof of address, checks them against the application and flags what a person needs to look at.
For claims, a decision model approves the simple, low-value claims straight away and sends the rest to a claims handler, with its reasons.
For customer contact, a fine-tuned small model reads each email or chat and routes it to the right team: the task we tested in section 2.
Lock-in builds up one reasonable decision at a time, until prices, laws or geopolitics change and there’s no room left to move.
Tied to one provider
Room to move
Built on what you control
Room to move
When prices, laws or geopolitics changeMove your data out and rebuild everything
Only the model changes
The foundation is cheapest to lay now, before your first hundred use cases are built on top of it.
Prove it works on one task you’d never send to a third-party LLM
Pick the low-hanging fruit: separate use cases, each with its own small model
Make AI a normal option, with each task on the model that suits its data
Revolut, Airbnb and bunq already run open models at a fraction of the cost per token, as long as the GPUs stay busy
A small model fine-tuned on our own tickets beat every model we tried, on a single GPU
How sensitive the data is decides where each task runs, from a frontier API to your own GPUs
| Model | Type | ECI | Release date |
|---|