Open-weights vs truly open source: the licensing debate surrounding AI models

In the rapidly evolving world of artificial intelligence, terms are thrown around constantly. One of the most controversial debates at the moment centers on the "open source" label. Large tech companies release powerful language models and label them as open source, while purists argue that these models are actually "open-weights". But what exactly is the difference, and why does it matter for developers and entrepreneurs?

What is Truly Open Source?

Historically, the definition of open source has been guarded by the Open Source Initiative (OSI). Software is open source if its license complies with the Open Source Definition (OSD). This means that anyone is free to use, modify, and distribute the software for any purpose, including commercial applications, without discrimination against specific individuals or groups.

The OSI is currently working on an official and definitive definition specifically for Open Source AI, which is expected to further standardize the dividing line in the industry [to be verified].

The Reality of Open-Weights

With "open-weights" models, the creator makes the resulting weights (parameters) of the trained neural network publicly available. This allows developers to run or fine-tune the model locally. However, unlike true open source, these licenses often impose restrictions.

Common restrictions in open-weights licenses include:

• Commercial limits: The model may not be used for free if your application or platform has more than a certain number of monthly active users.
• Acceptable Use Policies (AUP): Strict rules regarding the purposes for which the AI may or may not be used (e.g., prohibition of use in certain industries or for generating specific content).
• Prohibition on model training: The model may not be used to generate synthetic data to train a competing AI model.

Common Licenses at a Glance

To make the difference tangible, let's look at commonly used licenses in the AI ecosystem:

• Apache 2.0 / MIT: These are true, OSI-approved open source licenses. Models released under these licenses offer maximum freedom.
• Llama 3 Community License (Meta): A prime example of an open-weights license. It is widely accessible but includes a commercial threshold (700 million monthly users) and a prohibition on using the output to improve other language models.
• Gemma License (Google): Accessible to many developers, but requires agreement to specific terms of use that deviate from traditional open source principles.

Why This Debate Matters

For hobbyists, the semantics may matter very little; after all, they can download and test the models for free. But for entrepreneurs building applications, the difference is crucial. Building a business model on top of an open-weights foundation can lead to unexpected legal obstacles (vendor lock-in) when the application scales or when the license terms are modified in a subsequent model generation.

Are you looking for the right model for your next project and want to be sure of the freedoms you have? Consult our AI models directory for a categorized overview of current models and their exact licensing terms.