With open weights, a model's trained parameters are published. You can download it and run it on your own hardware — instead of only renting it through a provider's interface.
The weights are what training actually produces: billions of numbers holding what was learned. If they are open, you can take the model with you.
That is the decisive difference between renting and owning — and for many companies the point at which AI becomes approvable at all.
Open weights is not open source #
The two terms are constantly confused, and the difference is substantial:
- Open weights — the weights are available. Training data, training code and method usually stay secret. Licences often restrict certain uses.
- Open source in the strict sense — everything disclosed, including data and training process, under a recognised open-source licence. Few so-called "open-source models" meet that bar.
"Open" does not automatically mean "free to use". Some licences exclude commercial use above a certain size or require attribution. Check before production, not after.
When self-hosting pays off #
Running it yourself is more work and usually not cheaper. It pays off when one of these applies:
- Data must not leave the building. Health records, defence, legal, personnel files.
- Availability must be guaranteed. No provider can cut your access or change pricing overnight.
- The model must stay unchanged. With rented models the provider silently updates the version — what worked yesterday may answer differently today.
- Very high volume. Past a certain scale the cost calculation tips towards your own hardware.
If none of these applies, the rented service is almost always the sensible choice.
Why this matters in Europe #
Digital sovereignty stops being abstract the moment a company hangs its core processes on a model. Being able to run the model yourself makes you independent of one provider's decisions.
This is exactly where European providers position themselves: alongside commercial services they publish models with open weights, addressing organisations for which control over data and operations weighs more than the last point on a leaderboard.
Common questions #
What is the difference between open weights and open source?
With open weights only the trained parameters are available. Training data and code usually stay secret, and licence restrictions often apply. True open source discloses everything — few models called open source meet that bar.
Do I need my own hardware?
Smaller models run on a capable workstation. Large models need GPUs — in a business context, suitable infrastructure or a host that runs the model on your behalf.
Is an open model worse than a closed one?
Not inherently. The gap to the leaders has narrowed. What matters is whether the model solves your specific task well enough — not whether it tops a benchmark.
What does self-hosting really cost?
Hardware or hosting, operations, updates and people. Compare honestly against the rented service's usage fees. Below high volumes, self-hosting is rarely cheaper — the reason for it is control, not savings.
Sources #
- Regulation (EU) 2024/1689 (AI Act) — Art. 53 ff. and recitals on free and open-source AI models
- Open Source Initiative — Open Source AI Definition
- Mistral AI — publishes open-weight models alongside commercial ones
Rent or run it yourself?
Ask me — or ask Ada. She is my AI assistant, available around the clock, and will propose a slot with me if you need one. Free, no form.