The promise appeals: a model you install yourself, no subscription, no data leaving your walls. It is real, but the full calculation holds a few surprises, and it is better to know them before committing.
Definition
An open-source model is an AI model whose parameters are published, allowing you to download it and run it on your own hardware or at the host of your choice.
The term is slightly misleading. Most of these models publish their parameters but not their training data, and some licences restrict commercial use. Open-weight model is the more accurate description.
Always check the licence before building on it. Some prohibit commercial use, others allow it below a user threshold. A licence read after the fact can be expensive.
The real economics
| Closed model via API | Open model hosted | |
|---|---|---|
| Unit cost | Per use, by Token | Server paid continuously |
| Low volume | Very economical | Expensive: the server idles |
| High constant volume | The bill climbs | Becomes advantageous |
| Skills required | None | System administration |
| Confidentiality | Contractual | Technical |
The tipping point is higher than people imagine. A server capable of running a serious model costs several hundred euros a month, whether it works or not. It takes substantial, steady volume for the calculation to lean this way.
The two real reasons to go there
Technical confidentiality. When nothing must leave your infrastructure, no contractual commitment matches data that never travels. That is the case for health data and some regulated sectors.
Specialisation at volume. A small open model, Fine-tuning on your precise task, can match a large generalist model on that task at a far lower unit cost. Over millions of calls the gap becomes considerable.
Outside these two cases, a closed model's API remains almost always the rational choice, including for companies very attentive to their costs.
Do not forget the hidden cost: updates, monitoring, version upgrades, and the time of whoever looks after it. That adds to the server and is missing from most comparisons.
Frequently asked questions
Are open models worse?
The gap has narrowed markedly. On everyday writing or classification tasks the difference is often imperceptible. It widens again on long reasoning and tool use.
Can you try them without a server?
Yes, several hosts offer them as pay-per-use APIs. That is the best way to measure quality on your own cases before investing in a machine.
Is a laptop enough?
For small models and personal use, yes, and it is an excellent testing ground. For professional use with several people at once, no.
How do you know whether it is worth it in your case?
By first measuring your real token consumption over a month. The calculation becomes obvious once that figure is known, and our n8n course shows how to instrument that measurement in your scenarios.