Can an open-weight model be used commercially?

It depends on its license, which has to be read case by case: many allow it, but with restrictions an open source license would not carry, such as prohibited uses listed in an attached policy or thresholds beyond which a separate agreement with the provider is required. The availability of the weights file says nothing about usage rights.

Why are so many models called open source when they are not?

Because the term sells, and until a shared yardstick existed everyone used it their own way. Since October 2024 there is the Open Source Initiative's Open Source AI Definition, asking for parameters, code and information about the data: a good number of models presented as open source do not meet it, and the correct way to describe them is open weights.

Is running a model in house better than using an API?

It is better when the constraint is about data, meaning you must be able to state where inference runs and who can reach it, or when volumes are high and steady. For variable or experimental use, almost never: the API fee covers GPUs, updates and upkeep, which otherwise become your work and your infrastructure.

Are open weights and open source the same thing?

A model whose parameters you can download and run wherever you want, which makes it neither open source nor free of usage limits.

An open-weight model is one whose provider publishes the parameters: you can download them, run the model on your own infrastructure and adapt it with fine-tuning, without calling anyone's API. That is a substantial difference from a model reachable only as a service, and it bears directly on where your data ends up. It is not the same thing as open source, though, and that is the confusion marketing exploits most readily: weights are a file, not a license. That file almost always arrives with usage conditions, and almost never with the training data or the code that produced it. You can run the model, but you cannot answer the question of what went into it, which is exactly the question a customer will put to you in a supplier questionnaire, or an auditor during a review.

See also: DeepSeek · The Chinese language-model lab known for its Mixture-of-Experts architecture and open-weight releases.

What the reference definition says

The yardstick is not a matter of opinion. The Open Source AI Definition published by the Open Source Initiative in October 2024 asks for three things together: the parameters, the code used for training and inference, and enough information about the data that a third party could rebuild an equivalent system. Measured against it, a good number of models presented as open source are not, and the correct way to describe them is open weights. This is not lexical pedantry: it changes what you have the right to do and what you can state to whoever asks.

The conditions you read afterwards

Open-weight model licenses are often the provider's own, not standard software licenses, and they carry restrictions an open source license never would: prohibited uses listed in an attached policy, thresholds beyond which a separate agreement is required, attribution obligations, sometimes limits on what you may train from that model. They are worth reading before building on top, because the cost of finding out later is not a fee but a rewrite. It is also why an open-weight model reduces vendor lock-in without removing it: you move the dependency from the API to the license. The most widely used families in this category, from Llama to Qwen, carry different conditions and have to be compared one by one.

Why it matters for decision makers

The practical reason a small company looks at open weights is almost always one thing, and it is not cost: being able to say where inference runs and who sees the data, which is the first requirement of sovereignty and the first question from a regulated customer. The arithmetic has to be done in full, though, because running a model in house means GPUs, updates and somebody looking after them, while the API includes that work in the price. An honest comparison puts a subscription on one side and an infrastructure to maintain on the other, and for many uses the second does not pay for itself.