Open vs Closed Models: What 'Open Source AI' Really Means
Weights, data, licenses: the three things people conflate when they say open source.
When you ask a chatbot for a recipe, you're talking to a model. But what does it mean for that model to be "open source"? People use the phrase for three different things: the weights you can download, the data and code that built the model, and the license that says what you may do with them. Most models marketed as open are missing at least one of these pieces.
Here's the part that surprises most people: downloading the weights of a model like Llama does not make it open source. The Open Source Initiative (OSI), the group that maintains the definition of open source software, now has a formal definition for AI, and most "open" models do not meet it. Knowing which parts are actually open tells you whether you can audit a model, rebuild it, or put it in a product without a lawyer.
What does "open source AI" actually require?
The OSI's Open Source AI Definition, version 1.0, says an open source AI system must give you the freedom to use it for any purpose "without having to ask for permission", and to study, modify and share it (The Open Source AI Definition 1.0). To make that possible, three things must be available:
- Data information: detailed information about the data used to train the system, enough that "a skilled person can build a substantially equivalent system". The data itself is not always required, but the description must be that good.
- Code: the complete source code used to train and run the system.
- Parameters: the model's weights.
The code must come under an OSI-approved license, and the data information and weights under OSI-approved terms, meaning terms that grant the same freedoms.
This is a much higher bar than most people realize. When a company puts model weights on Hugging Face and calls them open source, it has almost always left out the training code and the detailed data information. What it released is open weights, not open source AI.
Why do people confuse weights with open source?
In traditional open source software, the source code is the thing. You download it, read it, change it and build it yourself.
Model weights are different. They are the output of training: billions of numbers that hold the patterns found in the training data. Having them lets you run the model and fine-tune it. It does not show you how the model was built, what data shaped it, or whether that data included copyrighted or biased material.
Think of a cake. Open weights are like getting a finished cake: you can eat it, slice it and add your own decoration. Open source AI is getting the recipe, the list of ingredients and where they came from, and the oven settings too, so you could bake a cake just like it yourself.
- Run the model yourself
- Fine-tune it on your own data
- Serve it behind your own app
- All of that, plus:
- Study and change the training code
- Rebuild a substantially equivalent model
- Audit what data shaped it
What's actually in a model release?
When a company releases a model, you usually get a few files. The weights file holds the billions of numbers the model learned. A config file describes its shape: how many layers, how wide, which kind of attention. A tokenizer file turns text into the tokens the model reads.
What you almost never get is the list of training data, the scripts that filtered and cleaned it, the exact training settings, or the tests used to measure the model. Those are the parts that would let someone rebuild or audit it. Without them, you can run the model but not fully understand it.
Some releases do go further. EleutherAI's Pythia models and the OLMo models from AI2 (the Allen Institute for AI) come with their training data and code. When the OSI tested its definition against real models in 2024, Pythia and OLMo were among those that met it, while Llama 2, Grok, Phi-2 and Mistral's Mixtral fell short (OSI's 2024 review of the definition).
How do licenses reveal what's really open?
The license is the fastest way to cut through the marketing. A permissive license like Apache 2.0 or MIT lets anyone use, change and share the weights, for any purpose, with no limit on who you are or how much you earn. If a model's license needs a separate deal once your company passes a size or revenue limit, it is not an open source license.
Even a permissive license on the weights is not enough by itself. Some Mistral models use Apache 2.0, an open source license for the weights, but without the training code and data information they are still open weights, not open source AI; Mixtral was one of the models that fell short in the OSI's tests. Other Mistral models, such as Devstral 2, use a modified MIT license that bars companies with more than $20 million in monthly revenue (Devstral 2 license). That kind of limit on who may use it is exactly what open source licenses forbid.
Meta's Llama license goes further. Among other terms, any company whose products had more than 700 million monthly active users on the release date must ask Meta for a license, which Meta may refuse (Llama 4 license). xAI shows how one company can do both: it released Grok-1 in March 2024 under Apache 2.0 (Grok-1 on GitHub), but the Grok 2 weights it released in August 2025 came with a custom license that forbids using them, or their outputs, to train other large AI models (Grok 2 on Hugging Face).
Open weights are now common, with every kind of license. DeepSeek releases its models under MIT, Alibaba's smaller Qwen models use Apache 2.0, and even OpenAI, best known for closed models, released its gpt-oss models under Apache 2.0 in August 2025 (gpt-oss-120b & gpt-oss-20b Model Card). What 'Open Source Model' Really Means: The Definition and the Licenses goes through these licenses in depth, with a table of popular models.
How do regulators think about this?
Regulators use their own categories, and they do not follow the OSI definition. The EU AI Act calls the big, broadly capable models "general-purpose AI models". Its Recital 98 gives an example: models "with at least a billion of parameters and trained with a large amount of data using self-supervision at scale" (Recital 98).
The Act does use the phrase "free and open-source licence". The provider of a general-purpose model released under such a license, with its weights and information about its design and use made public, is excused from two duties: the detailed technical documentation and the information for companies that build on the model. It must still have a copyright policy and publish a summary of its training content, and the exemption does not apply to models with "systemic risk", the most powerful ones (Article 53). The test is about the license and public weights, not about OSI's full definition.
In the United States, the 2023 executive order on AI was revoked in January 2025 (Executive Order 14148). The government's AI Action Plan of July 2025 has a section called "Encourage Open-Source and Open-Weight AI", which describes such models as ones "made freely available by developers for anyone in the world to download and modify" (America's AI Action Plan). As in the EU, the focus is on whether the weights are available, not on the OSI's full definition. So a model can count as "open" for a regulator while failing the OSI's test.
Why would anyone choose closed?
With a closed model, you use it through the provider's app or API, and the provider handles everything else: the computers that run it, the safety filters, the updates. You have no GPUs to buy, no serving software to maintain and no weight files to protect. Many of the most capable models are closed, too.
The provider can also fix problems centrally. When a new way of tricking models appears, it can patch its service once for everyone, while people running open weights must update their own copies.
The trade-off is dependency. Everything you send goes to the provider, so you rely on its contract and privacy terms, and it controls the model's abilities, price and availability. If it retires a model or changes its rules, you adapt or move.
Why would anyone choose open weights?
Privacy. You can run the model on your own computers, so your data never leaves them. The US AI Action Plan makes the same point: many businesses and governments "have sensitive data that they cannot send to closed model vendors".
Offline and local use. An open model can run on a laptop or a private server with no internet connection. Local Inference: Running Models on Your Own Machine shows what that takes.
No per-token bill. You pay for the computers, not for each token. At high volume that can be cheaper, and at low volume it can be more expensive.
No forced upgrades. The model you tested stays exactly the same for as long as you keep it.
Freedom to adapt. You can fine-tune it, within its license, for your own task.
The trade-off is work: you must run it, secure it, keep it updated and check its quality yourself.
Quick Reference
| Property | Open source AI | Open weights | Closed model |
|---|---|---|---|
| Weights you can download | Yes | Yes | No |
| Training code | Yes | Usually no | No |
| Data information | Detailed | Minimal or none | Minimal or none |
| License | OSI-approved terms | From permissive to custom and restricted | Provider's terms of service |
| Can rebuild a substantially equivalent model | Yes, in principle | No | No |
| Can fine-tune | Yes | Usually yes, within the license | Only where the provider offers it, which is now rare |
| Can deploy commercially | Yes | Depends on the license | Through the provider's service |
| Examples (as of October 2026) | Pythia (EleutherAI), OLMo (AI2) | Llama, DeepSeek, Qwen, gpt-oss, Mistral | GPT, Claude, Gemini |
Frequently Asked Questions
Q: Can I use Llama in a commercial product without paying Meta? Usually yes, under the Llama license, which allows commercial use with conditions: among them, companies above 700 million monthly active users need a separate license, and there are naming and acceptable use rules. It is not an OSI-approved license, so read the specific terms for your case.
Q: What's the practical difference between Apache 2.0 weights and a custom license? Apache 2.0 gives you lasting rights to use, change and share the weights for any purpose, with only a copy of the license and notices to pass on. A custom license can add conditions such as revenue limits, banned uses or rules about training other models. Read the grant clause before you build.
Q: Does "open weights" mean I can see what data the model trained on? No. Open weights means you have the final trained numbers. Information about the training data is separate, and most open-weight releases give only a short description of their sources. Without more, you cannot fully check for copyrighted material or bias.
Q: Are there any truly open source large language models? Yes, but they are rare. EleutherAI's Pythia and AI2's OLMo release weights, training code and training data, and both met the OSI's definition in its 2024 tests. Most other models marketed as open source are open weights.
Q: How does the EU AI Act affect my choice between open and closed models? It depends on your role. The duties for general-purpose AI models fall on providers, the companies that develop or release the model, and open-weight releases under a free and open-source license get some exemptions. If you only build an app on someone else's model, different duties apply to you, depending on how you use it, such as telling people they are talking to an AI. Check the obligations for your situation.
Test yourself
You're choosing a model for a customer support chatbot that will handle sensitive financial questions. Model A is closed: you use it through an API, and the provider offers security certifications and a contract that limits what it may do with your data. Model B is called "open source" in its marketing: its weights are downloadable under a custom license that forbids use in financial services without a separate agreement, and its training data is described only as "a mix of public web data and licensed sources". Which model is the safer choice for your compliance team, and why?
Answer: Model A, despite being closed. Its provider makes written commitments about how your data is handled and holds certifications you can show to auditors and regulators. Model B is not open source at all: its license restricts a field of use, which open source licenses forbid, and it explicitly blocks financial services, so you would need a separate agreement anyway. Its vague data description also means you cannot check what it learned from, a question your compliance team will ask. If you wanted an open-weight option, look for a model whose license permits financial use, such as Apache 2.0, ideally with detailed data information, and run it on your own servers so the sensitive data never leaves them.
Next: Fine-Tuning, Explained: Teaching an Old Model New Tricks. It shows what changes inside a model when you train it further on your own examples, and when you should not bother.
If you want this kind of breakdown every week, how real AI systems actually work under the hood, subscribe to Internals Decoded at internalsdecoded.com.
Sources
- The Open Source AI Definition 1.0 (Open Source Initiative)
- 2024 end-of-year review: Open Source AI Definition v1.0 (Open Source Initiative)
- Devstral 2 license (Mistral AI, read October 2026)
- Llama 4 license (Meta)
- Grok-1 (xAI, March 2024)
- Grok 2 (xAI, August 2025), README and license
- gpt-oss-120b & gpt-oss-20b Model Card (OpenAI 2025)
- DeepSeek-V4.1-Flash model card (DeepSeek, October 2026), MIT license
- Qwen3.8-27B model card (Alibaba, August 2026), Apache 2.0
- EU AI Act, Recital 98 and Article 53
- Executive Order 14148, Initial Rescissions of Harmful Executive Orders and Actions (January 2025)
- America's AI Action Plan (The White House, July 2025)
