The difference in one sentence
A closed model is reachable only through an API or an application. It runs on the vendor's servers, and what you buy is the service.
An open-weight model has weights you can download and run wherever you like: on a company server, on a rented machine, or even on your own laptop.
What "open" does not mean here
This is worth clearing up, because the word misleads.
Open-weight is not the same as open source. The weights are downloadable, but the training data almost never is, and neither is the full training procedure. You cannot reproduce the model, only run it.
And it is not necessarily free to use as you please. Licences differ: some restrict commercial use, some charge a fee above a revenue threshold. Restrictions are more common on the largest and strongest open-weight models than on the smaller ones. You have to read the licence before you use one — this is not a formality.
The three criteria that actually decide
1. Where does your data sit?
This is the most common genuine reason to go open-weight.
With a closed model the data leaves the company. Business tiers from serious providers do contract not to train on it, and EU data residency is available — in many cases that is entirely sufficient.
But there are places where it is not: health data, legal privilege, a contractual obligation that forbids passing data to a third party. In those cases open-weight is not the better choice, it is the only choice.
One caveat, though: running it on your own server is not data protection in itself. Configure it badly and you are worse off than you would have been with a serious provider.
2. What does it cost?
Here a widespread misconception needs clearing up: open-weight is not free. The model is free; running it is not.
With a closed model you pay per token. There is no base fee, no operations overhead, and zero requests cost zero.
With open-weight you pay for hardware — bought or rented — plus operations. That is a fixed cost, and it runs whether anyone uses it that day or not.
The arithmetic follows straightforwardly: at low and medium volume the closed model is cheaper. At very high, sustained volume the balance tips, because the fixed cost gets spread out. The crossover point differs by company, but for most small and mid-sized businesses it is a long way off.
3. What happens if the vendor changes something?
This criterion is the least glamorous and, in hindsight, the most painful.
With a closed model the vendor can retire the model, alter its behaviour or raise the price at any time. This is not theoretical: model retirements are a routine event, and when one happens the prompts you tuned to it will behave differently.
A downloaded open-weight model, by contrast, stays exactly the same. If it works today, it will work identically in five years. In a regulated environment, where a validated process cannot simply be changed, that is a decisive argument.
As for capability
It used to be simple: closed models were stronger. Today it is more nuanced.
The strongest models are still closed, but the gap has narrowed noticeably, and the best open-weight models now sit at the level the closed frontier occupied not so long ago. For the majority of tasks that is plenty.
The practical question is not which is stronger but whether it is good enough for your task. You can settle that in half a day with a test set of twenty real examples — far more reliably than from any leaderboard. That is what How to Read AI Benchmarks is about.
A decision order
Do not start from which camp appeals to you. Ask in sequence:
- Is there data that legally or contractually cannot leave the company? If yes, open-weight, self-hosted. The remaining questions are only about implementation.
- Will the volume be very high and predictable? If yes, cost out both options. Self-hosting may well pay.
- Neither? Start with a closed model. You get to a working solution faster, and you can switch later at any time.
The third case is the most common one, and there is no shame in it. The worst outcome is spending six months building infrastructure before you find out whether the task can be solved with AI at all.
What to take away
The choice is not a creed but three questions: where the data is allowed to be, how large the volume is, and how much it hurts if the vendor changes something. Answer all three and the decision makes itself.