What 'Open Source Model' Really Means: The Definition and the Licenses

You met the basics in Open vs Closed Models: What 'Open Source AI' Really Means: downloadable weights are not the same as open source. Here we go deeper: what the Open Source Initiative's Open Source AI Definition 1.0 (OSAID) actually asks for, how to read the licenses of today's popular models, and how to classify a release yourself.

The short version: most models marketed as open source, including Meta's Llama family, do not meet the OSI's criteria. They are downloadable and you can adapt them heavily (fine-tune them, distill them, merge them), but restrictive licenses, missing training code, and undocumented training data make them impossible to reproduce or fully audit. That 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 cannot sell it if the bakery's terms say you cannot.

Diagram loading

What exactly does the OSI definition require?

The Open Source AI Definition starts from four freedoms: use the system for any purpose without asking permission, study how it works, modify it for any purpose, and share it, with or without changes. To make those freedoms real for a machine learning system, it requires three things, in the form a person would actually use to make changes:

• Data information: "sufficiently detailed information about the data used to train the system" so that a skilled person can build a substantially equivalent system.
• Code: "the complete source code used to train and run the system", including the code that processed and filtered the data, the training code, the inference code and the model architecture.
• Parameters: "the model parameters, such as weights or other configuration settings."

The code must be under OSI-approved licenses; the data information and the parameters must be shared under OSI-approved terms. Just posting a checkpoint file somewhere does not cut it.

The definition does not demand that the raw training data be published in full. That would be legally impossible for many datasets containing copyrighted material or personally identifiable information (PII). Instead the provider must give a complete description of all the data used, including any data that cannot be shared, plus a list of all publicly available training data and where to obtain it, and of all data obtainable from third parties, including for a fee. This keeps the door open for reproducibility without forcing data owners to break privacy laws.

The freedoms rule out restrictive licenses. A license that makes some companies ask permission, bans some uses, or forbids using outputs to train other models fails the first freedom. The OSI has said so explicitly about Meta's licenses, first for Llama 2 and then for Llama 3.x.

Diagram loading

Why are most "open" models not actually open source?

They release weights and inference code, but they omit the training code, the data information, 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 is the textbook case. You can download the weights, run inference, and fine-tune. But the licenses for Llama 3.1 and Llama 4 add conditions that open source licenses do not have (Llama 3.1 license, Llama 4 license):

• Any company whose products had more than 700 million monthly active users (MAU) on the release date must request a license from Meta, which Meta may grant or refuse.
• You must display "Built with Llama", and any model you train or improve with Llama or its outputs must have "Llama" at the start of its name.
• You must follow Meta's acceptable use policy. For Llama 4, that policy also withholds the rights to its multimodal models from individuals and companies based in the European Union, though not from end users of products built on them.

Earlier licenses went further: Llama 2 and the first Llama 3 release forbade using the model's outputs to improve any other large language model. As for data, the Llama 3.1 model card describes its training data only as about 15 trillion tokens "from publicly available sources", with no training code. Critics call this openwashing: using the "open" label to borrow goodwill while keeping the system closed in the ways that matter for reproducibility and auditability.

The RAIL family (Responsible AI Licenses) adds another wrinkle. These licenses grant distribution and modification rights but attach use restrictions that must be passed on to everyone downstream (OpenRAIL on Hugging Face). Stable Diffusion's original CreativeML OpenRAIL-M license, for example, bans uses such as discriminating against people. Well-intentioned or not, such terms fail the definition's "any purpose" freedom, and the RAIL initiative's own FAQ says plainly that these are not open source licenses. The original Stable Diffusion repository (its code and 1.x weights) and the SDXL weights use RAIL licenses, and Stable Diffusion 3.5 uses Stability AI's own Community License, which allows commercial use for free only to organizations and individuals with under 1 million dollars in annual revenue.

What is the difference between closed, open-weight and fully open?

Three tiers cover almost every release:

• Closed: you get an API (application programming interface) only, as with GPT-4 or Claude. You never see weights, code, or data. You rent access.
• Open-weight: you get a downloadable checkpoint, usually with inference code, under some license.
• Fully open: weights, training code, detailed data information, and open licenses.

The practical difference decides whether you can reproduce training, audit the system, or fork it without legal headaches.

AI2's OLMo models are built to be fully open. The OLMo 2 paper, playfully titled 2 OLMo 2 Furious, releases model weights, full training data, training code and recipes, training logs and thousands of intermediate checkpoints, and it groups other models into open-weight, partially open and fully open. Its successor, Olmo 3 (November 2025), again releases every component, with the code and models under Apache 2.0 and its Dolma 3 training data under the ODC-By license. That lets a well-resourced team reproduce the model from scratch, something impossible with Llama.

Permissive open-weight releases are now common. xAI published Grok-1 in March 2024, a 314-billion-parameter Mixture-of-Experts model (one split into many expert sub-networks, of which only a few run for each token), as weights and example code under Apache 2.0 (xAI's announcement). You can use those weights commercially without field-of-use restrictions, but the training data and training code were never released. Its successor release is different. Grok 2, the model xAI used in 2024, was published as weights in August 2025 (news reports called it Grok 2.5) under a custom license that forbids using the model or its outputs to train other foundation or large language models (Grok 2 on Hugging Face). The license was released as the Grok 2 Community License and has been titled the xAI Community License since November 2025. OpenAI's gpt-oss models (August 2025) come under Apache 2.0, but OpenAI released the weights, inference code, tool environments and tokenizer, not the training code or data (gpt-oss-120b & gpt-oss-20b Model Card). Qwen3 (Apache 2.0) and DeepSeek-R1 (MIT) are other permissive open-weight families. A permissive license makes them easy to build on; it does not by itself make them open source.

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 OSAID'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 show the model was not trained on copyrighted material or sensitive personal data.

EleutherAI's story shows why provenance matters. It published The Pile: An 800GB Dataset of Diverse Text for Language Modeling, an 825 GiB English corpus, openly. But one of its parts, Books3, was a collection of pirated books, and it was taken offline after an anti-piracy group sent a takedown notice in August 2023 (TorrentFreak's report). In June 2025 EleutherAI released the Common Pile, 8 TB of openly licensed and public domain text from 30 sources, together with Comma, a pair of 7B models trained on it. Openly published data is not automatically clean data; documented, licensed data is what lets others build on it with confidence.

Even when raw data cannot be shared, the OSAID's data information requirement (how the data was collected, labeled, and filtered, and where the public parts can be found) gives you enough to attempt a reproduction. Without it, you have no path to equivalence.

What licenses qualify as open source for AI?

For the code, only OSI-approved licenses; for the weights and data information, terms that grant the same freedoms. 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.

For an engineer deciding whether to build on a model, the license is the contract. An Apache 2.0 release like Grok-1's or gpt-oss's weights gives broad rights for commercial use and modification. A custom license like Llama's or Grok 2's requires legal review and may block your plans. The label "open source" is not a substitute for reading the actual terms.

Quick Reference: Openness Spectrum at a Glance

Facts as of October 2026, from each project's own release pages linked above.

Model	Weights download	Training code	Data information	License	Open source under the OSAID?
Olmo 3 (AI2)	Yes	Yes	Yes, data released	Apache 2.0 (data ODC-By)	Built to be; OSI's own checks passed the earlier OLMo
Llama 3.1 and Llama 4 (Meta)	Yes	No	One-line summary	Custom Llama license	No
gpt-oss (OpenAI)	Yes	No (inference code only)	No	Apache 2.0	No
Grok-1 (xAI)	Yes	No (example code only)	No	Apache 2.0	No
Grok 2 (xAI)	Yes	No	No	xAI Community License (released as the Grok 2 Community License)	No
GPT-4, Claude	No	No	No	API terms	No

Frequently Asked Questions

Q: Is Meta's Llama open source? No. It is an open-weight model under a custom license with a 700 million monthly active user clause, naming and branding rules, and an acceptable use policy, and Meta does not release its training code or detailed data information. The OSI has said Meta's licenses are not open source.

Q: Can I fine-tune an open-weight model and redistribute my fine-tune? Depends on the license. Llama's license allows it if you pass on the license, display "Built with Llama", and start your model's name with "Llama". Apache 2.0 and MIT models allow it with only a copy of the license and notices. Other licenses may restrict redistribution or downstream use. Always check the specific terms.

Q: Why does the OSI allow withholding raw training data? Raw data often contains copyrighted material or personal data that cannot be lawfully redistributed. The OSI requires instead a complete description of the data, plus where to get the public and third-party parts, 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, and detailed data information) in the form you would use to modify them. A permissively licensed checkpoint is open-weight, not fully open.

Q: Is Stable Diffusion an open source model? No. Its weights are downloadable, but the original release (code and 1.x weights) and the SDXL weights use RAIL licenses with use restrictions, and Stable Diffusion 3.5 uses Stability AI's Community License with a revenue limit. Neither kind is an open source license.

Test yourself

A new model ships its weights under Apache 2.0 with inference code and a one-paragraph summary of its training data ("a mix of web, code and books"), but no training code. Is it open source under the OSAID? Can your company sell a product built on it?

Answer: It is not open source under the OSAID. A one-paragraph summary is not data information detailed enough for a skilled person to build a substantially equivalent system: it does not describe how the data was collected, filtered and labeled, or list where the public parts can be obtained. And without the training code, the "code" requirement fails too. It is a permissively licensed open-weight model. Selling a product on it is a separate question, and the answer is yes: Apache 2.0 allows commercial use, modification and redistribution, as long as you keep the license and its notices with what you distribute. Before you ship, still check whether the release adds a separate usage policy or trademark rules, as some do.

If you want this kind of breakdown every week, how real systems actually work under the hood, subscribe to Internals Decoded at internalsdecoded.com.

Sources

• The Open Source AI Definition 1.0 (Open Source Initiative)
• Meta's LLaMa 2 license is not Open Source (Open Source Initiative)
• Meta's Llama license is still not Open Source (Open Source Initiative)
• Llama 3.1 license (Meta, GitHub)
• Llama 4 license (Meta, GitHub)
• OpenRAIL licenses explained (Hugging Face blog)
• RAIL licenses FAQ (Responsible AI Licenses)
• 2 OLMo 2 Furious (AI2, 2025), the OLMo 2 paper
• Olmo 3 announcement (AI2, November 2025)
• Open release of Grok-1 (xAI, March 2024)
• Grok 2 model card and license (xAI on Hugging Face)
• gpt-oss-120b & gpt-oss-20b Model Card (OpenAI, 2025)
• The Pile: An 800GB Dataset of Diverse Text for Language Modeling (Gao et al. 2020)
• Anti-piracy group takes prominent AI training dataset Books3 offline (TorrentFreak, August 2023)
• The Common Pile v0.1 (EleutherAI, June 2025)