Understanding the Distinction Between Open Weights and Open Source

In the evolving landscape of Artificial Intelligence, there seems to be a common misconception regarding what constitutes “open source.” Many individuals refer to various downloadable models as open source simply because they can download and run the model weights locally. However, this interpretation overlooks a crucial aspect of software development and accessibility.

When we speak of downloadable model files, such as .gguf or .safetensors, we must recognize that these files are akin to executable files (.exe) in traditional software. In essence, they represent “compiled AI” rather than the underlying open-source framework itself. The true essence of an open-source project includes not only the model weights but also the source code associated with the framework utilized to train and deploy the model—think of frameworks like Llama and Mistral—as well as the datasets that provided the foundational training.

Unfortunately, this is where many offerings fall short. As it stands, very few major AI providers disclose the actual source code, which includes the datasets utilized in training their models. While many projects are labeled as open source, they do not meet the full criteria necessary for that designation. OASST is one notable exception, but even high-profile projects, often referred to as open source, such as DeepSeek, do not fit the bill entirely.

It is important for the community to understand the implications of this distinction. A genuinely open-source AI model, complete with publicly accessible training datasets that anyone with sufficient computing power could use to “recompile” or modify from scratch, could transform the AI landscape in a manner that parallels the revolutionary impact of the Linux kernel in the operating system domain.

As we move forward in AI development, let’s advocate for transparency and a true understanding of what open source means, ensuring that the community is better informed about the tools at their disposal.