1 · What changed, and why it matters 2 · The proof: our own test 3 · What it means for you

ADC · November 2026

From one model
to a portfolio

Open-weight models on your own infrastructure

Sound familiar?

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

Every organisation has dozens of tasks where AI could help

Today: one model for everything, or no AI at all

The future is a portfolio: the right model for each task

AI today
No AI yet

One frontier model provider via API

Data too sensitive

Not good enough, or too costly

Not implemented yet

Frontier model

Complex reasoning

Model A

Bulk, non-sensitive work

Model B

Sensitive data, self-hosted

Model C

High-volume, narrow tasks

Model D

Fast classification

The why, the proof and the how

1

What changed, and why it matters

2

The proof: our own test

3

What it means for you

Two terms to know

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.

Open weights

Download the model, run it anywhere

Small language models

Up to 100× smaller and cheaper than frontier models

The capability race

Most AI solutions don’t need the very best model. More than fifteen open models from 2026 score in this range.

Capability score, Epoch Capabilities Index
The frontier

Closed models keep raising the bar

Open weights

Open models follow close behind

The gap today

~4 months

Why move to open models now?

1

Sovereignty

  • Your data.
  • Your model.
2

Business case

Cost
Latency
Quality

Sovereignty is already being tested

June 2026

The model can be switched off

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.

June 2025

US law follows the data

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.

June 2025

Your provider can drop you

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.

Sovereignty is a scale, not a binary

1Chinese API

Data to China, where by law companies must help state intelligence

2US API

Data to the US, where authorities can demand it, within limits the EU accepts

3US cloud, EU region

Data stays in the EU, but US law can still make the provider hand it over

4EU provider API

Data and provider both under EU law

5Your EU cloud

Your own model, on GPUs you rent in the EU

6Your own GPUs

Model, data and hardware are all yours

Closer to you, more sovereign

Many organisations today

Open models are already cutting token bills

October 2026

AI coding costs down 90%

About 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.

November 2026

Frontier quality at 1% of the cost

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.

October 2026

Support answers 40% cheaper

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 tried it on a real task

We ran open and closed models on support tickets, sorting each one by intent, team and priority.

Customer ticket

LLM
Routing decision

      
Four ways we ran it, from least to most sovereign
1US API
2EU cloud, rented GPUs
3Our own GPUs
4Our own GPUs, fine-tuned

Fine-tuned: trained further on our own labelled tickets

Fine-tuning teaches a small model a specific task

You can only do this when you have the weights, which is a big reason open weights matter.

Start

Open-weight model

Billions of weights, like these sliders

+
Add

Your labelled examples

→
Result

A specialist on your task

Now it answers the way your team would

What you need

Fewer than 1,000 labelled examples from your history, and one GPU to train on

When it’s worth it

Narrow, frequent tasks, where the same question comes up thousands of times

The catch

When the task changes, you have to train the model again

Your own GPUs pay off when they’re busy

Illustrative shape from published break-even analyses. Where the lines cross depends on your model, hardware and prices.

Paying per token

With an API, cost rises with every token

Your own GPU

A fixed cost, busy or idle

Break-even

Past this point, extra tokens cost almost nothing

Against a cheap API

A hosted open model moves break-even further right

Scaling up

Cost rises in steps: each new GPU starts out idle

Plan for utilisation: share GPUs across use cases and fill quiet hours with batch work.

What we learned

Our results match what public benchmarks and other teams have already found.

No trade-off on quality

Out of the box, the best open model did the job about as well as the best closed model we tried.

Fine-tuning on your own data is the edge

Fine-tuned on our own past tickets, a small open model beat every model we tried, the frontier included.

Fast, on hardware you can own

The fine-tuned model ran on one GPU and answered several times faster than the frontier.

Six places to start

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.

Secure assistant

Open weights, self-hosted

AML alert triage

Fine-tuned small model

Coding agent

Open weights, self-hosted

KYC document checks

Open weights, self-hosted

Claims: approve or escalate

Decision model

Customer contact routing

Fine-tuned small model

What you build today determines how resilient you are in the future

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

Quick start on their API
Prompts tuned to their model
Workflows built on their features
Your data in their platform

Room to move

Built on what you control

One open model on your own infrastructure
A layer that lets you swap models
Your own test sets to compare models
Data and fine-tunes in formats you own

Room to move

When prices, laws or geopolitics change

Move 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.

How to get to the right model for every task

Phase 1

Prove it works on one task you’d never send to a third-party LLM

1 sensitive task
1 open-weight model
1 local GPU, like a DGX Spark

Phase 2

Pick the low-hanging fruit: separate use cases, each with its own small model

A handful of separate use cases
Small fine-tuned models
Desk-side GPU workstations

Phase 3

Make AI a normal option, with each task on the model that suits its data

Tasks in every team
A portfolio of models
Shared GPU servers that grow with use

From one pilot to a handful

  • Pick the low-hanging fruitSeparate use cases with clear value
  • Reuse the recipeOpen model, your own examples, a test set
  • Add desk-side GPUsOne workstation per use case
  • A small core teamPeople who can fine-tune and test models
Big step

The organisational shift

  • InfrastructureFrom desk-side GPUs to GPU servers: fully on premises, or with an EU host
  • ProcurementBuy and own your GPUs, or contract EU hosting and models on standard terms
  • Shared services, open to every teamOne platform: models, fine-tuning and support for any team
  • Governance and complianceA full governance framework, with a clear route to adoption across the organisation

Three things to remember

1

Open models cut the token bill where volume is high

Revolut, Airbnb and bunq already run open models at a fraction of the cost per token, as long as the GPUs stay busy

2

A small model fine-tuned on your data can beat the frontier

A small model fine-tuned on our own tickets beat every model we tried, on a single GPU

3

Sovereignty is a choice per task, not all or nothing

How sensitive the data is decides where each task runs, from a frontier API to your own GPUs

Your AI, on your own terms

Solution Advisor, Benelux
Click or press → to continue
1
Record-setting models on the Epoch Capabilities Index, January 2023 to October 2026
ModelTypeECIRelease date