In the AI gold rush, founders are building on models described as 'open'. But the terms 'open weights' and 'open source' are not the same, and misunderstanding the difference is a business risk that can derail a startup before it scales.
The 'Open Source' We Know and Trust
For decades, 'open
source' has had a clear meaning in the software world. It means the source code is publicly available for anyone to use, study, modify, and distribute. Governed by permissive licenses like MIT or Apache 2.0, open-source software became the bedrock of the modern internet, allowing developers to build on a shared foundation without fear of restrictive legal terms. This model fosters collaboration and innovation. When a project is truly open source, you have the freedom to build a commercial product on it, customize it to your needs, and inspect it for security flaws. The Open Source Initiative (OSI) acts as the official steward, providing a clear definition that prohibits licenses from discriminating against fields of use or persons.
The New Contender: 'Open Weights'
The term 'open weights' is a newer concept specific to AI. It means a model's developer has made its trained parameters—the 'weights' that hold all the model's learned knowledge—available for download. This allows anyone with the right hardware to run the model themselves, fine-tune it with their own data, and gain a degree of control impossible with closed, API-only models like OpenAI's most advanced systems. However, 'open weights' says nothing about your right to use the model commercially or how it was made. Unlike true open source, the training data and training code are often kept secret. The model is a product you can run, not a transparent system you can fully rebuild or audit.
The Fight Is in the Fine Print
The core of the fight comes down to the license attached to the weights. While some open-weight models are released under permissive, OSI-approved licenses like Apache 2.0, many of the most popular models are not. Industry giants have released powerful models like Llama and Gemma under custom 'community licenses'. These licenses often look permissive at first glance but contain crucial restrictions. For example, a license might prohibit commercial use once your product reaches a certain number of monthly active users or your company exceeds a revenue threshold, at which point you need a separate, often costly, commercial agreement. Others include 'acceptable use' policies that can restrict use in certain industries, which you are then contractually obligated to enforce on your own customers.
A Minefield for Founders
For a founder, betting your startup on an open-weight model without scrutinizing the license is a strategic gamble. If you build a successful product on a model with a revenue or user cap, your own growth could trigger a licensing crisis, forcing you into an expensive negotiation or a complete architectural rebuild. Enterprise customers may also balk during procurement if they discover your tech stack is built on a model with restrictive terms, seeing it as a compliance risk they don't want to inherit. Furthermore, if you fine tune a model with your own proprietary data, the model's original license might attach to your new, derivative work, creating IP contamination that can complicate fundraising or a future acquisition. These are not edge cases; they are fundamental business risks that stem directly from the difference between truly open source and merely open weights.
