Open-weight or paid AI models: what is the difference?

Some recipes are kept secret, and you can only taste the dish at the restaurant. Others are published, so you can cook them in your own kitchen. You still need the kitchen, the ingredients and someone who can cook.

AI models split along a similar line. Some are offered only as a paid service. Others publish their weights so anyone can run them. This piece explains what "open" means for a model, what it does not mean, and why the difference matters most for who runs the machines.

What the weights are

A trained model is, at its core, a very large file of numbers called weights. The weights hold what the model learned in training. Running the model means loading those weights onto suitable hardware and sending it text.

Whoever holds the weights can run the model. Whoever does not can only reach it through someone else. We explain the difference between training and running in Training vs inference.

Paid models: access through a service

With a paid, closed model, the maker keeps the weights. You send requests over the internet to their service and pay per use, usually per token.

What you get. No hardware to manage, quick start, and a provider who handles updates and scaling.

What you give up. You cannot see or change the weights. The model runs where the provider chooses, on their terms, and they can change or retire it.

Open-weight models: the file is yours to run

With an open-weight model, the maker publishes the weights for download. You, or a provider you choose, can run the model on any suitable hardware.

What you get. Choice of where it runs, which can help with data residency and latency. The option to fine-tune it for your own task. A model version that does not change unless you change it.

What you take on. Someone has to supply the hardware, run the software, monitor it and keep it secure. Our piece on what it takes to run an open model yourself looks at this.

What "open" does not mean

The word "open" is used loosely, so it helps to be precise.

Not always open source. Publishing weights is not the same as publishing the training code, the data or the full method. Many open-weight models share only the weights.

Not always free to use. Licences vary. Some allow broad commercial use. Others restrict certain uses, set conditions above a certain size of user base, or require attribution.

Not free to run. The download may cost nothing, but the machines, power and people to run it do. Open weights move the cost; they do not remove it.

Not automatically safer or riskier. Security depends on how a model is deployed and monitored, not only on whether its weights are public.

Because licences differ, read the terms for each model before relying on it. This article is general information, not legal advice.

Why it changes who runs the machines

The main practical difference is not price per token. It is who operates the hardware.

With a paid model, the maker or its chosen cloud runs everything. With an open-weight model, the work can run on many kinds of infrastructure: your own servers, a large cloud provider, or a specialist inference operator.

That opens choices a closed model cannot offer. You can pick a location close to your users. You can keep data in a particular country. You can move the workload if terms or costs change. We cover location rules in Data residency explained.

Questions to ask when choosing

Neither path is right for every workload. Before deciding, it helps to ask:

1. Does the task need a very capable model, or would a smaller open model do? Our piece on when a smaller model is enough covers this.
2. Where must the data be processed, and can a paid service meet that?
3. What does the licence allow for my use?
4. Who will run, monitor and update the model if we choose open weights?
5. What happens to my application if the model is changed or retired?
6. How do total costs compare, including the costs that are not on the quote?

For dated prices and speeds across both kinds of model, see Model Watch.

What we are building

Open-weight models make the question of where AI runs a practical one. They need operators with the right hardware in the right places.

We're building standardized compute close to power, connectivity and demand, connected through one programmable network for AI inference. Workloads are designed to be placed by policy, on latency, data residency, model and load, and the network is designed so every workload can generate an auditable record of commitment, placement and delivery.

We are building for a world where you can choose the model and also choose where it runs.

We're building towards running open-weight models without you operating the hardware yourself. If that is what you are planning, use Request capacity below.