Many artificial intelligence models marketed as "open source" are actually only "open-weight"—a distinction with real consequences for enterprises deploying the technology.
Models such as Llama, DeepSeek, Qwen, and Gemma allow users to download the finished model and run it on their own hardware. This enables fine-tuning on proprietary data and removes reliance on external APIs, offering cost control and vendor independence.
But open weights come without the underlying recipe. The training code, exact datasets, and filtering processes remain proprietary. Users get the finished parameters—not the means to rebuild or audit them.
A truly open-source model includes both the trained weights and everything needed to reproduce it: training code, data specifications, and sufficient documentation to audit the model from scratch. The distinction is not semantic. It determines what users can legally do, whether internal composition can be trusted, and what "open" actually means.
The Open Source Initiative defines open-source AI as requiring materials and legal freedom to use, study, modify, and share the system for any purpose. The organization publicly pushed back against Meta's 2024 framing of Llama as open source, highlighting what it called "openwashing" in the industry.
Developers working with open-weight systems cannot change or even examine the training data. A developer cannot remove specific datasets—say, Reddit data—to create a safer model. This limits verification of and control over core components.
Companies deploying open-weight models must trust the creator's choices on data and methodology. That opacity creates governance and regulatory compliance risks in business applications.
The competitive advantage lies in control of the factory itself. Labs that release only weights retain proprietary knowledge of the creation process, sustaining strategic advantages even as model access widens.
