Examples
Worked examples
- Is an instance
Llama-3-70B released under the Llama 3 Community Licence with downloadable weights.
- Is an instance
Mistral-7B released under Apache-2.0 with downloadable weights.
Counter-examples
Looks similar, but isn't
- Not an instance
A model accessible only via paid API.
- Not an instance
An open-source AI system meeting the OSI's full data+code+weights definition (a stricter superset).
Editorial commentary
An open-weights model is one whose trained parameter values (the weights file) are published for download and local use, without necessarily disclosing the training data, training code, or full methodology needed to reproduce the model from scratch. Llama, Mistral, Qwen, DeepSeek, Mixtral and Gemma are the models most commonly described this way. Publishing weights alone is enough to enable on-premises deployment, fine-tuning, and inspection of model behaviour that an API-only model does not allow — which is why open-weights releases matter operationally for research computing and procurement even where the underlying training process stays undisclosed.
This label is contested, not settled. Many of these same releases are marketed by their developers as “open source,” and the term is used inconsistently across the field. The Open Source Initiative’s Open Source AI Definition (OSAID v1.0, October 2024) draws a hard line: to qualify as genuinely open source under OSAID, a release must give downstream users enough information about training data, training code, and model architecture/parameters to substantially recreate an equivalent system — not just a downloadable weights file. Measured against that bar, OSI’s own analysis places Llama, Gemma and Qwen in the open-weights category, not open source, because their training data is not disclosed to that standard. See open-source model (criteria) for the OSAID side of this dispute in full, including why the disagreement matters for procurement and licence review rather than treating it as a semantic quibble.
Licensing detail that follows from this
Open-weights releases typically carry a non-OSI-approved custom licence rather than a standard open-source licence. The Llama Community License (3.x), for example, grants free commercial and research use but prohibits using Llama’s outputs to train competing large language models, and terminates the free grant if the licensee’s products exceed 700 million monthly active users, at which point a separate licence must be negotiated with Meta. Research offices reviewing an “open” model for procurement or reuse should read the actual licence text rather than relying on the open/open-source label its developer applies to it.
References
- Open Source Initiative, Open Source AI Definition v1.0 (October 2024)
- Meta, Llama Community License (3.x), llama.com
Also known as
open weights · open-weight model
Machine-readable encodings
Use in your systems
<role vocab="credit"
vocab-identifier="https://casrai.org/dictionary/"
vocab-term="Open weights model"
vocab-term-identifier="https://casrai.org/dictionary/term/open-weights-model" />{
"@context": "https://schema.org",
"@type": "DefinedTerm",
"@id": "https://casrai.org/dictionary/term/open-weights-model",
"name": "Open weights model",
"identifier": "https://casrai.org/dictionary/term/open-weights-model",
"description": "A model whose trained parameter values are publicly released and downloadable, typically under a named licence, distinct from but often described as 'open' even when training data and code are not released.",
"inDefinedTermSet": "https://casrai.org/dictionary/domain/ai-ml-research-outputs#set",
"url": "https://casrai.org/dictionary/term/open-weights-model",
"sameAs": [
"open weights",
"open-weight model"
],
"license": "https://creativecommons.org/licenses/by/4.0/",
"publisher": {
"@id": "https://casrai.org/#organization"
},
"author": {
"@id": "https://casrai.org/#editorial-team"
},
"datePublished": "2026-05-21T02:22:51",
"dateModified": "2026-08-22T15:54:49",
"inLanguage": "en-GB",
"isAccessibleForFree": true
}
