In plain terms
Renting a furnished flat or buying a house. The closed model is the flat: you move in today and the landlord fixes what breaks, but the landlord can also change the terms or renovate while you live there. The open model is the house: it is yours to alter and nobody can take it away, and the roof, the plumbing and the insurance are now your problem.
Why it matters
Six questions decide it. Control: must data and model stay in your hands, and the model unchanged? Cost: closed is pay-per-use with no fixed cost; open is cheap per request only at steady volume. Capability: the strongest models are usually closed. Support: a provider brings service levels and safeguards; with open models that work is yours or a hosting partner's. Licence: “open” covers everything from unrestricted to non-commercial. Risk: a closed supplier can raise prices or retire a model; open weights bring operational and security duties. Most organisations end up using both.
Example
A logistics company compares two options for reading 2 million shipping documents a year. A closed model through an API: 96% field accuracy, 70,000 dollars a year, nothing to operate. An open-weight model on rented GPUs: 93% after fine-tuning, 45,000 dollars for capacity plus half an engineer. It chooses the closed model for the documents and, for a separate project on drivers' health records, an open model on its own servers.
Most often confused with
Open source vs. open weights
When buying, the label matters less than the licence. Open-source AI in the full sense, with training code and data details under an unrestricted licence, is rare. Most models called open are open-weight, and their terms range from Apache 2.0, which permits almost everything, to custom licences that limit commercial use, scale or field. Ask for the licence by name and have legal read it before a model enters a product.
Under the hood
The two-way label hides three delivery forms: a closed model through the provider's API or a cloud platform; an open-weight model run by a hosting provider and billed per token, which needs no operations; and an open-weight model you host yourself. Compare them on an evaluation set from your own task; public benchmarks rarely settle it. Cost: closed is variable cost per token; self-hosted is mostly fixed cost and depends on utilisation; hosted open models sit in between. Customisation: closed models offer prompting and, for some, managed fine-tuning; open weights allow full fine-tuning, LoRA adapters, quantisation and inspection of internals. Stability: an API model is deprecated on the provider's schedule, while downloaded weights never change. Licences: Apache 2.0 and MIT are permissive; other families use their own licences with conditions, such as Meta's Llama licence; the Open Source Initiative's 2024 definition of open-source AI also asks for training code and detailed information about the data. A gateway and a model-independent test set keep a later switch cheap in either direction.