What 'Open Source Model' Really Means

Calling a model “open source” because you can download its weights is misleading. True open source AI, as defined by the Open Source Initiative’s Open Source AI Definition 1.0, requires far more: training code, inference code, and enough detail about the training data that a skilled person can build a substantially equivalent system. Most releases today are open-weight, not open source.

The vast majority of models marketed as open source, including Meta’s LLaMA family, do not meet the OSI’s criteria. They are downloadable, but restrictive licenses, missing training code, and opaque data pipelines make them impossible to reproduce or meaningfully modify beyond surface-level fine-tuning. This gap has real legal and engineering consequences.

Think of a model release like a bakery. Open source software gives you the recipe, the ingredient labels, and the right to modify and sell the result. An open source AI model, in the OSI’s framing, is the full recipe book: the training code, a detailed sourcing guide for the data, the inference code, and the weights. An open-weight release is more like a finished cake delivered to your door. You can eat it, slice it, maybe add frosting. But you cannot figure out how to bake it yourself, and you definitely cannot sell it if the bakery’s terms say you cannot.

How does the OSI define an open source AI model?

The OSI’s Open Source AI Definition 1.0 treats an AI system as a composite of structural elements and requires the “preferred form to make modifications” for each one. You get the four classical open source freedoms: use, study, modify, and share.

The definition explicitly names the base model, the trained model, weights and parameters, training code, inference code, and documentation of the training process as structural elements. For an AI model to be open source, the release must include the data information and code used to derive the parameters. Just posting a checkpoint file somewhere does not cut it. The OSI insists that a “skilled person” must be able to build a substantially equivalent system from what you provide. That means the training recipe and data description are non-negotiable.

Critically, the definition does not mandate that raw training data be published in full. That would be legally impossible for many datasets containing copyrighted material or personally identifiable information. Instead, the provider must disclose a complete description of the data: provenance, scope, characteristics, collection and selection methods, labeling procedures, and filtering. Lists of public datasets with locations and instructions for obtaining third-party datasets are required. IBM analysis This carve-out keeps the door open for reproducibility without forcing data owners to violate privacy laws.

A license that qualifies must also grant all four freedoms without discrimination against persons, groups, or fields of endeavor. This immediately disqualifies licenses that ban commercial use above a certain scale or forbid using outputs to train other models. OSI blog

Why are not most “open” models actually open source?

They release weights and inference code, but they omit the training code, the data recipe, or both. Often they come with custom licenses that restrict what you can build or who can use them. This pattern is called open-weight, and it fails the OSI test.

Meta’s LLaMA 3 is a textbook case. You can download the weights, run inference, and fine-tune. But the license prohibits commercial use by organizations with over 700 million monthly active users, forbids improving other models with LLaMA outputs, and imposes naming rules on derivatives. Meta license That is clear field-of-use discrimination. The training code and data details are absent. Critics rightly label this as openwashing: using the “open” label to borrow goodwill while keeping the system closed in the ways that matter for reproducibility and auditability.

The RAIL license family (Responsible AI License) introduces another wrinkle. It grants distribution and modification rights but adds contractual use-case restrictions. For example, a RAIL license might ban discriminatory applications or development of weapons. While well-intentioned, these terms violate the OSI’s non-discrimination principle. Models under RAIL, such as some Stable Diffusion variants, look open on the surface but are not open source by the definition.

What is the difference between open weights, open access, and fully open?

Open weights gives you a downloadable checkpoint. Open access is API-only. Fully open means weights, training code, detailed data information, and an OSI-approved license. The practical difference determines whether you can reproduce training, audit the system, or fork it without legal headaches.

The OLMo 2 project from AI2 deliberately spans the gap. It releases weights, data, code, intermediate checkpoints, and instruction-tuned variants under Apache 2.0. The authors explicitly classify it as “fully open” and distinguish open-weight (weights only) from partially open (weights plus some data details) and fully open releases. OLMo 2 That approach lets a well-resourced team reproduce the model from scratch, something impossible with LLaMA.

Grok-1 sits in a middle ground. xAI published the 314B Mixture-of-Experts weights and architecture under Apache 2.0, so you can use the weights commercially without field-of-use restrictions. Grok-1 repo The training data pipeline, however, is undocumented. You cannot reproduce training. It is open-weight with a permissive license, not fully open.

Closed models, GPT-4, Claude, only expose an API (application programming interface). You never see weights, code, or data. The vendor owns everything. You rent access.

How does training data affect what you can do with a model?

Without detailed data information, you cannot reproduce the model, audit bias, or verify legal compliance. The OSI’s data requirement is the bridge from a black box to a system you can reason about.

If you are deploying a model in a regulated industry, you may need to show what data was used for training. An open-weight release without data provenance leaves you in the dark. You cannot prove the model was not trained on copyrighted material or sensitive personal data. The EleutherAI project deliberately built “The Pile” as a public 800 GB corpus so that others could build on it with confidence. EleutherAI interview That explicit commitment to data openness is a key differentiator between fully open and open-weight.

Even when raw data cannot be shared, the OSAID’s data information clause, describing how data was collected, labeled, and filtered, gives you enough to attempt a reproduction. Without it, you have no path to equivalence.

What licenses qualify as open source for AI?

Only licenses that meet the OSI’s open source definition. That means no field-of-use restrictions, no discriminatory terms, and the right to modify and redistribute. Apache 2.0, MIT, and BSD are the typical choices.

Hugging Face’s licensing guide draws a sharp line between “publicly available” and “open source.” A model you can download but that comes with a restrictive license is not open source. RAIL licenses, despite permissive base terms, add behavioral restrictions that make them non-OSI. So while they may be useful for some communities, they are not open source in the formal sense.

For an engineer deciding whether to build on a model, the license is the contract. An Apache 2.0 release like Grok-1’s weights gives broad rights for commercial use and modification. A custom license like LLaMA’s requires legal review and may block your monetization plan. The label “open source” is not a substitute for reading the actual terms.

Quick Reference: Openness Spectrum at a Glance

Frequently Asked Questions

Q: Is Meta’s LLaMA open source? No. It is an open-weight model under a custom license that restricts commercial use and derivative model training. It fails the OSI’s definition on license terms and missing training artifacts.

Q: Can I fine-tune an open-weight model and redistribute my fine-tune? Depends on the license. LLaMA’s license allows redistribution with attribution but places limits on downstream use. Other open-weight models may forbid redistribution entirely. Always check the specific terms.

Q: Why does the OSI allow withholding raw training data? Raw data often contains copyrighted material or PII that cannot be lawfully redistributed. The OSI requires instead a detailed description sufficient for a skilled practitioner to build a substantially equivalent system.

Q: Does an Apache 2.0 license on weights make a model open source? Not on its own. The OSI definition demands the full set of artifacts, training code, inference code, detailed data info, in the preferred form for modification. A permissively licensed checkpoint is open-weight, not fully open.

Q: Is Stable Diffusion an open source model? The base code is open under a permissive license, and weights are downloadable. But training data was not fully documented, and some variants use RAIL licenses with use-case restrictions. It is partially open, not OSI-compliant.

If you want this kind of transparency breakdown every week, mapping real system internals, licensing pitfalls, and what “open” actually means, subscribe to Internals Decoded at internalsdecoded.com.

Sources

• IBM: Open Source AI Definition
• Jisc: Open-source AI, a guide for decision-makers
• Orange: Open-source AI: A typology of AI models
• OLMo 2: Fully Open Language Model
• Meta LLaMA 3 Community License
• EleutherAI Year One
• Grok-1 GitHub
• Stable Diffusion GitHub
• opensource.org · Blog