I’ve been writing about open source for over 25 years and there is always some form of controversy to talk and write about.
In 2025 that controversy is all about open source AI — whatever that term means.
Outside of AI, open source to me is a very specific thing. It’s not a marketing label (though it is frequently used as such by vendors large and small), it is a specific license and approach to development. Simply put, when it comes to (non-AI) software, open source is software is licensed under an Open Source Initiative (OSI) compliant license. OSI is the keeper of the open source definition, so that makes sense to me.
But that doesn’t work for AI (apparently).
With AI, large vendors (I’m looking at you Meta) built seemingly open models (Llama) but didn’t use an OSI license and then called the model open source, creating all kinds of confusion. To make matters worse, the OSI has put out a new definition of what open source AI means, that few people —myself included — agree with.
The challenge with an AI model is that it’s not just the code for the model, but also the model weights as well as the data that make a model work. Without all three, there is only a degree of openness. DeepSeek, for example, is available under the OSI-approved license, making it somewhat open (and technically open source).
The benefits of true open source has always been the ability to use code freely and build your own innovations on top of it. You can do that with DeepSeek, you can do that with Llama too (albeit with restrictions), you can’t do that (easily) with OpenAI.
So what exactly is “open source” when it comes to AI? Out in the quickly in evolving AI sector, that confusion remains….
