What open weights means
A trained language model is, at its core, a large file of numbers called parameters or weights. Open weights means that file is published and anyone can download it and run the model on their own hardware. It does not say under which license, whether the code used to train the model is available, or what data went into it. Those are separate questions, and they are where the labels split.
Running a model yourself changes the economics compared with paying per token for a hosted model. The AI cost comparison shows what the same job costs on hosted models from their official prices, and the token counter measures how many tokens your prompts and documents have. The guide running AI models locally covers the memory side.
The Open Source AI Definition 1.0
The Open Source Initiative (OSI) released version 1.0 of the Open Source AI Definition on October 28, 2024, at the All Things Open conference in Raleigh, North Carolina. It asks that an AI system grant four freedoms: Use, Study, Modify and Share. In the definition's words, the first is to use the system for any purpose and without having to ask for permission; the others are to examine how it works, change it for any purpose, and share it with or without changes.
Exercising those freedoms needs what the definition calls the preferred form to make modifications. For a machine-learning system it must include 3 elements:
| Element | What must be available | Under |
|---|---|---|
| Data information | Enough detail about the training data for a skilled person to build a substantially equivalent system | OSI-approved terms |
| Code | The complete source code used to train and run the system | OSI-approved licenses |
| Parameters | The model parameters, such as weights or other configuration settings | OSI-approved terms |
Two details are often misread. First, the training data itself does not have to be published. What is required is a full description of it, including how it was obtained and processed, and a list of where public and third-party data can be obtained. Second, the definition allows license conditions that require modified versions to be released under the same terms as the original, the way copyleft licenses do for software.
Where software open source comes in
The older Open Source Definition for software has ten criteria, from free redistribution and access to source code to a technology-neutral license. Two of them matter most for model licenses. Criterion 5 says the license must not discriminate against any person or group of persons. Criterion 6 says it must not restrict anyone from making use of the program in a specific field of endeavor. Licenses that limit who may use a model, or for what, set conditions that these two criteria do not allow for software.
Three kinds of release
Downloadable models fall roughly into three groups by what is published and under which terms:
| Release | Weights | Code | Data | Terms |
|---|---|---|---|---|
| Weights, custom license | Yes | Varies | Varies | Own conditions |
| Weights, open license | Yes | Varies | Varies | e.g. Apache 2.0 |
| Full release | Yes | Training and running | Data or full description | Open terms throughout |
Only the last group can meet the AI definition, and only if every element is under open terms. The first two are open weights. Whether a specific model qualifies is a question about its license and its documentation, not its name.
Examples from real licenses
Three releases, described from their own license texts and announcements:
- A custom community license. The Llama 3.1 Community License Agreement, with a version release date of July 23, 2024, lets you download and use the models with conditions. Products with more than 700,000,000 monthly active users in the preceding calendar month must request a license from Meta, which it may grant at its sole discretion. Products built on the models must display Built with Llama. Use must follow an acceptable use policy that is part of the agreement. And a model trained or improved with Llama outputs and then distributed must have Llama at the beginning of its name.
- A standard open source license on the weights. Mistral AI announced Mistral 7B, a 7.3 billion parameter model, on September 27, 2023, under the Apache 2.0 license, and wrote that it can be used without restrictions.
- A full release. Ai2 describes its Olmo models as a fully open language model and complete model flow. Next to the weights it links the pretraining, mid-training and post-training datasets and the tools used to train the models.
All three are open weights. They differ in what you may do with the model and in what you can learn about how it was made.
When the difference matters
- Building a product. Read the license for user thresholds, required notices, acceptable use terms and naming rules before you ship.
- Sharing a fine-tuned model. Some licenses carry conditions into derivatives, such as naming rules or the requirement to release under the same terms.
- Research and audits. Studying how a model behaves is easier with the training code and a description of the data. Weights alone show what the model does, not how it got there.
- Running it at home. Even for private use, check the license once; after that the question is whether the model fits in your memory, which the guide on running models locally works through.
The labels matter because they compress all of this into one word. When a model is called open, check which of the three elements are published and read the license. This page describes the definition and the licenses; it is not legal advice. For how models read text in the first place, see what is a token in AI.
Frequently asked questions
What does open weights mean?
The trained parameters of a model, the weights, can be downloaded and run on your own hardware. The label says nothing about the license, the training code or the training data. Some open-weight models come under standard open source licenses, others under custom licenses with their own conditions.
Is an open-weights model open source?
Not automatically. Under the Open Source AI Definition 1.0, an open source AI system must also provide the complete code to train and run it and detailed information about its training data, all under OSI-approved terms, and must allow use for any purpose without asking permission. Weights alone cover only one of the three required elements.
Does open source AI require releasing the training data?
Not the data itself. The definition requires sufficiently detailed information about the data, including how it was obtained and processed and where public or third-party data can be obtained, so that a skilled person could recreate a substantially equivalent system.
Can I use an open-weights model in a commercial product?
It depends on the license, so read it. Apache 2.0 places few conditions on use. A custom license can add conditions, for example the Llama 3.1 Community License requires a separate license from Meta above 700 million monthly active users, a Built with Llama notice, and compliance with an acceptable use policy. This page describes licenses; it is not legal advice.
Why does the difference matter if I only want to run the model?
For running a model at home it often does not. It matters when you build on the model, share a fine-tuned version, use it in a product, or want to know what it was trained on. Those are exactly the rights and materials that differ between an open-weights release and an open source one.
Sources
- Open Source Initiative, The Open Source AI Definition 1.0 (the freedoms to use, study, modify and share; preferred form with data information, code and parameters under OSI-approved terms; data information instead of the data itself; conditions that require modified versions under the same terms)
- Open Source Initiative, announcement of the Open Source AI Definition 1.0, October 28, 2024, at All Things Open in Raleigh
- Open Source Initiative, The Open Source Definition (the ten criteria; 5, no discrimination against persons or groups; 6, no discrimination against fields of endeavor)
- Meta, Llama 3.1 Community License Agreement, version release date July 23, 2024 (700 million monthly active users clause, Built with Llama, Acceptable Use Policy, Llama at the beginning of derived model names)
- Mistral AI, Announcing Mistral 7B, September 27, 2023 (7.3B parameters, released under the Apache 2.0 license, usable without restrictions)
- Ai2, Olmo (a fully open language model and complete model flow, with linked pretraining, mid-training and post-training data and training tools)