Open Weights vs Truly Open: Licenses, Made Simple

Artificial intelligence (AI) models are often described with one word: open. That word can mean very different things. Sometimes it means you can download the model weights. Sometimes it means you can use the model commercially. Sometimes it means researchers can inspect how it was trained. A license is the document that turns those possibilities into permissions, limits, and duties. If you learn to read the few clauses that matter, the fog clears quickly.

Start with what is actually being shared

A model is not a single thing in the way a book or a photograph is a single thing. The part people talk about most is the weights: the learned numbers that let the model turn inputs into outputs. If you can download the weights, you can often run the model yourself, adapt it, or study its behavior.

But weights are only one part of the story. A release may also include inference code, training code, tokenizer files, evaluation results, safety notes, documentation, and sometimes information about the training data. Each of these can have its own copyright status, license, or usage terms.

So the first question is not, Is the model open? The first question is, What exactly did I receive, and under what terms?

Weights: the learned numerical parameters of the model.
Code: software used to run, train, or adapt the model.
Data information: details about what the model learned from, even when the raw data is not shared.
Documentation: instructions, limits, evaluation results, and safety notes.
License: the permission document that says what you may do with the release.

A model release is several parts, each with its own terms and its own level of transparency.

Open weights are access, not freedom

Open weights means the weights are available to download. That is valuable. It can reduce dependency on a hosted service, make local testing possible, and let engineers tune or optimize the model for a specific environment.

But access is not the same as freedom. A model can have open weights and still forbid commercial use, redistribution, certain industries, certain countries, or use by organizations above a revenue threshold. Another model can have downloadable weights under a permissive license that allows broad use and sharing. Both may be called open in casual speech, but they are not equivalent.

Think of openness as a spectrum. At one end, you can only call a hosted system. At the other, you have the materials and legal permission needed to inspect, use, modify, and share the model without discrimination against people, groups, or fields of work.

Openness runs from a hosted service you can only call to a model you may inspect, modify, and share.

What a license actually does

A license is permission in writing. Without permission, the default rule in many legal systems is simple: you do not automatically have the right to copy, modify, distribute, or build a business on someone else’s copyrighted work. A license changes that default by granting rights, usually with conditions.

Model licenses often combine several ideas. Copyright controls copying and modification of protected materials such as code, documentation, and sometimes model files. Contract terms may add promises you accept when downloading or using the model. Patent terms matter when the model or its methods may be covered by patents. Trademark terms matter when you want to use a project or company name in your own product.

This is why two licenses that look similar at first glance can create very different outcomes. One may grant broad copyright rights but no patent protection. Another may allow use but only if you follow an acceptable-use policy. A third may allow internal testing but not redistribution to customers.

⚠️

Not legal advice. Use this guide to ask better questions, and involve counsel before betting a product, dataset, or customer contract on a license interpretation.

The clauses that change the answer

You do not need to memorize every license. Start by looking for the clauses that decide what your team can and cannot do. These clauses usually determine whether a model can move from experiment to product.

Commercial use is the first fork in the road. A noncommercial license may be fine for learning, research, or a proof of concept, but it can block paid products, internal business workflows, or customer-facing features. Redistribution is the next fork. If you cannot redistribute the model or a modified version, you may still be able to use it internally, but you may not be able to ship it with software, place it in a customer environment, or publish a fine-tuned version.

Modification rights matter when you want to fine-tune, quantize, prune, merge, or otherwise adapt a model. Attribution terms matter when you must preserve notices or credit the original authors. Share-alike or copyleft terms matter when your changes must be released under the same license. Acceptable-use terms matter when the license bans particular uses, such as deception, surveillance, weapons, or regulated decisions.

Can we use it commercially, including inside the business and in paid products?
Can we modify it, fine-tune it, compress it, or combine it with other models?
Can we redistribute the original model, our modified model, or an application that includes it?
Must we publish changes, preserve notices, or use the same license for derivatives?
Does the license restrict fields of use, users, geography, revenue, or deployment scale?
Is there a patent grant, and does it end if we bring a patent claim?
Are there separate terms for outputs, data, benchmarks, names, or logos?

The license clauses decide whether a model can move from experiment to shipped product.

Open source rules do not bend just because it is AI

Open source has a specific meaning in software. The Open Source Initiative (OSI) has long used rules that reject discrimination against people, groups, and fields of endeavor. In plain English: a license is not open source if it says you may use the software only for research, only if you are a small company, or only outside certain industries.

AI complicates the package, but not the core principle. A model license can include safety restrictions that many people consider responsible, useful, or necessary. But if those restrictions limit who may use the model or what field it may be used in, the license is usually not open source in the classic OSI sense.

That does not make the model bad. It makes the label important. A model can be responsible, useful, downloadable, and still not be open source. Calling it open weights is often more accurate.

📌

Useful is not identical to open. A restrictive license may be the right choice for a project, but it should not be described as fully open.

Truly open needs more than permission to download

In this article, truly open means more than downloadable weights. It means a capable person can inspect, use, modify, and share the model, and can understand how it was made well enough to evaluate its strengths, limits, and risks. That requires legal permission and practical transparency.

The hardest part is often training data. Many model builders cannot release the full dataset because it contains licensed material, personal information, trade secrets, or data gathered under terms that do not allow redistribution. They may still publish a data mixture, filtering method, source categories, or a datasheet that helps others reason about the model.

The same is true for the training recipe. Batch sizes, optimization settings, compute budget, filtering rules, evaluation prompts, and safety tuning can materially affect the final model. If those details are missing, others may be allowed to use the model but unable to reproduce or properly audit it.

How leaders and practitioners should decide

The real decision is not whether a model has an attractive label. The real decision is whether the model’s rights, obligations, and evidence match the job you need it to do. A research notebook, an internal assistant, an embedded customer product, and a regulated workflow may require different answers.

For a practitioner, the license affects architecture. If redistribution is forbidden, you may need to keep the model on your own servers instead of shipping it to customers. If commercial use is unclear, you may need a different model before launch. If derivatives must use the same license, you need to know whether your fine-tuned model, adapters, or merged weights are derivatives under the terms you accepted.

For a leader, the license affects risk, procurement, and exit options. You need to know whether your team can keep using the model if your company grows, whether customer contracts allow the model’s terms, whether outputs can be stored or sold, and whether the license creates obligations your compliance process can actually meet. The same analysis applies whether the model is downloaded from a repository, selected in Microsoft Foundry, or consumed through an application programming interface (API).

Name the exact model, version, files, and source location.
Write down the intended uses: research, internal production, customer product, redistribution, or fine-tuning.
Check commercial use, modification, redistribution, attribution, share-alike, acceptable-use, and patent terms.
Record whether the license applies to weights, code, documentation, outputs, and derivatives.
Decide who owns ongoing monitoring if the model, license, or deployment use changes.

⚠️

Do not approve the label. Approve the specific license, version, model files, and intended use, not a web page that simply says open.

A simple rule of thumb

If you remember one sentence, make it this: open weights tell you that you can get the model; the license tells you what you can do with it; true openness tells you whether others can understand, modify, and share it in a meaningful way.

That sentence prevents most confusion. It separates access from rights, rights from transparency, and transparency from trust. A model with open weights may be perfect for your use case. A truly open model may be better for auditability, long-term independence, or community improvement. A hosted closed model may still be the right business choice. The point is to choose with clear terms instead of slogans.

Key takeaways

Open weights means the model weights are downloadable; it does not automatically mean commercial use, modification, or redistribution is allowed.
A license is the practical source of permission, and small clauses can decide whether a model can move from experiment to product.
Classic open source rules reject discrimination against people, groups, and fields of use; many AI model licenses with safety or commercial restrictions are better described as open weights.
True openness combines broad legal rights with practical transparency about data, training, evaluation, and limitations.
Leaders should approve a specific model version, license, and intended use, not a vague open label.

Glossary

Weights: Weights are the learned numerical values inside a model that determine how it responds to inputs.
License: A license is a written permission that says what you may do with someone else’s model, code, documentation, or other materials.
Open weights: Open weights means the model weights are available to download, even if the license limits what you may do with them.
Open source: Open source describes a license that grants broad rights to use, study, modify, and share software without discrimination against people, groups, or fields of work.
Truly open: Truly open describes a model release that provides broad rights and enough supporting information to inspect, adapt, and meaningfully understand the model.
Permissive license: A permissive license allows broad use, modification, and redistribution with relatively few obligations, often mainly notice and attribution.
Copyleft: Copyleft is a licensing approach that allows sharing and modification but requires some downstream versions to use the same or compatible license.
Field-of-use restriction: A field-of-use restriction is a license term that forbids use in certain activities, industries, or domains.
Redistribution: Redistribution is giving copies of the original or modified model materials to someone else.