What Makes a Language Model "Open"? Weights, Data, Code and Licences
"Open" has become one of the most contested words in language technology. A model can be called open because anyone can download it, because its code is public, or because every step from raw data to final weights can be inspected and repeated. These are very different promises. This guide separates them.
What is actually in a model release?
When a lab "releases a model", it may publish any combination of the following:
| Component | What it lets you do |
|---|---|
| Model weights | Run the model and fine-tune it on your own data |
| Inference code | Load the weights and generate output correctly |
| Training code and configuration | See how the model was trained; repeat or vary the process |
| Training data, or a full description of it | Check for contamination, bias, personal data and copyright issues |
| Intermediate checkpoints and logs | Study how capabilities emerge during training |
| Documentation (model card, data statement) | Understand intended uses, limitations and evaluation |
| Licence | Know what you are legally allowed to do with all of the above |
A useful way to think about this, proposed by Irene Solaiman in "The Gradient of Generative AI Release" (FAccT 2023), is as a spectrum from fully closed (access only through a company's interface) to fully open. Researchers at Radboud University led by Andreas Liesenfeld and Mark Dingemanse, in "Opening up ChatGPT" (ACM Conversational User Interfaces 2023), scored instruction-tuned models on more than a dozen separate openness dimensions and found that many models marketed as open disclosed little about their training data and fine-tuning.
Open weights are not open source
In software, "open source" has had a precise meaning since the late 1990s: a licence approved by the Open Source Initiative (OSI), which must allow anyone to use, study, modify and redistribute the code for any purpose, without discrimination against people or fields of use. Many downloadable language models do not meet this standard:
- Meta's Llama 2 licence (July 2023) allowed broad use but required companies whose products had more than 700 million monthly active users to request a separate licence, and attached an acceptable use policy. The OSI stated publicly that this licence was not open source.
- The BigScience workshop released BLOOM in July 2022, a model trained on 46 natural languages and 13 programming languages, under a "Responsible AI License" (RAIL) that permits free use and redistribution but forbids a list of harmful uses. Such behavioural restrictions are a deliberate ethical choice, but they make a licence incompatible with the open-source definition, which forbids field-of-use restrictions.
- Some models, by contrast, are published under standard permissive software licences: Mistral AI released Mistral 7B under the Apache 2.0 licence in September 2023.
Even an Apache-licensed set of weights may come with no information about training data. Because weights alone cannot be audited for what went into them, many researchers prefer the term "open weights" for such releases. Licences for the data side are covered in our guide to dataset licences for NLP.
The Open Source AI Definition
After a public, multi-stakeholder consultation, the OSI released version 1.0 of the Open Source AI Definition on 28 October 2024. It extends the four software freedoms (use, study, modify, share) to AI systems and lists what must be available under OSI-approved terms:
- Data information: sufficiently detailed information about the training data for a skilled person to build a substantially equivalent system, including its provenance, scope, selection, labelling and processing.
- Code: the complete source code used to train and run the system, including data processing and filtering.
- Parameters: the model weights and other configuration settings.
The definition does not require the training data itself to be published, which drew criticism from some open-data advocates who argued that a model cannot really be studied without its data. The OSI's position is that some data (personal or copyrighted material, for instance) cannot legally be shared, and that detailed information is the realistic minimum.
Fully open research models
A small number of projects publish everything. EleutherAI's Pythia suite (2023) released a series of models of different sizes trained on the same public data in the same order, with intermediate checkpoints, specifically to support research on how models learn. The Allen Institute for AI's OLMo, released in February 2024, published weights, training code, training logs and its pretraining corpus, Dolma. Such releases make it possible to check claims, trace a model's behaviour back to its data, and repeat experiments, the subject of our article on reproducibility in NLP.
What the EU AI Act says
The EU's Artificial Intelligence Act, Regulation (EU) 2024/1689, entered into force on 1 August 2024, and its obligations for providers of general-purpose AI models apply from 2 August 2025. Article 53 requires these providers to keep technical documentation, give information to downstream developers, put in place a policy to comply with EU copyright law, and publish a sufficiently detailed summary of the content used for training. Article 53(2) exempts models released under a free and open-source licence, with their parameters and architecture publicly available, from the two documentation duties, but not from the copyright policy or the training-content summary, and not if the model presents "systemic risk" as defined by the Act. In other words, EU law also treats openness as a matter of degree, and still expects open models to say what they were trained on.
How to judge a release
- Read the licence, not the announcement. Look for user thresholds, use restrictions, and terms that forbid using outputs to train other models.
- Look for the data. Is the training corpus published, or described in enough detail to rebuild it? Is there a datasheet or data statement (see datasheets and model cards)?
- Check for training code and checkpoints if you need to study or reproduce the model, not only use it.
- Check the evaluation: are test sets and prompts published so that you can rerun them?
- Record versions. Weights hosted online can be updated; note the exact release you used.
Sources
- Open Source Initiative, The Open Source AI Definition, version 1.0, 28 October 2024 (opensource.org).
- Regulation (EU) 2024/1689 (Artificial Intelligence Act), Articles 53 and 113, Official Journal of the EU, 12 July 2024.
- Solaiman, "The Gradient of Generative AI Release: Methods and Considerations", ACM FAccT 2023.
- Liesenfeld, Lopez and Dingemanse, "Opening up ChatGPT: Tracking openness, transparency, and accountability in instruction-tuned text generators", ACM CUI 2023.
- BigScience Workshop, "BLOOM: A 176B-Parameter Open-Access Multilingual Language Model", 2022.
- Biderman and colleagues, "Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling", ICML 2023.
- Groeneveld and colleagues, "OLMo: Accelerating the Science of Language Models", ACL 2024.
Open researchDataset licences for NLP
CC BY, ShareAlike, NonCommercial, ODbL, LDC agreements and the EU's text and data mining exceptions, explained for people who build and share corpora.
Open researchReproducibility in NLP
Preprocessing, seeds, versions and silent API updates: why NLP results fail to repeat, what studies found, and a routine that makes your work checkable.
Open researchDatasheets, data statements and model cards
The three documentation formats every NLP researcher should know, what each asks, and how they connect to EU transparency rules.