Open-Weight vs. Open-Source AI: What 'Open' Really Means

"Open" has become one of the most debated words in AI. Not every model you can download is open source in the traditional sense.

Open weights

An open-weight model makes its trained parameters (weights) available to download and run. This lets organizations host models themselves, fine-tune them and keep data in-house.

Licenses differ

• Some models are released under permissive licenses such as Apache 2.0, allowing broad commercial use.
• Others use custom licenses that may restrict certain uses or very large-scale deployments.
• Training data and code are often not released, even when weights are.

The Open Source AI Definition

In 2024 the Open Source Initiative (OSI) published the Open Source AI Definition, describing the freedoms — to use, study, modify and share — and the information about training data and code needed for an AI system to qualify as open source.

What teams should check

• The exact license and any usage restrictions
• Whether commercial use and fine-tuning are allowed
• What documentation exists about training data and evaluation

Key takeaways

• Open weights enable self-hosting and customization.
• "Open" licenses vary — read them before you build.
• Open-source AI standards are still evolving.