NVIDIA this week threw its weight behind an effort to create a standard for a licensing framework for artificial intelligence (AI) models that is being led by the Linux Foundation.
With the release of version 1.1 of the OpenMDW licensing framework, NVIDIA has decided to use this licensing framework across its families of Cosmos, Isaac GR00T, Ising and Nemotron models.
That endorsement should hopefully convince other providers of open source models to soon follow suit, says Mike Dolan, senior vice president of legal and strategic programs for the Linux Foundation.
Currently, most providers of open source AI models are relying on MIT or Apache licensing frameworks, but they are not designed from the ground up to address how AI models are built and deployed, says Dolan.
The OpenMDW licensing framework, in contrast, addresses the entire AI model, including weights, documentation, training data and output rather than just the software the model is based on.
Designed to be compatible with the Model Openness Framework, the overall goal is to create a standard licensing framework that provides a common definition for what it means for an AI model to be open source, says Dolan.
In effect, the Linux Foundation is committing to do the boring stuff that ensures organizations don’t find themselves, for one reason or another, locked into a specific AI model, he adds. “We’re kind of the janitors of open source,” says Dolan.
Ultimately, it will be up to each organization that opts to use a so-called open AI model to understand the caveats. In some cases, the software might only be available under an open source license but the weights used to create it remain proprietary. In other circumstances, there may be dependencies on components that may not have permissive open source licenses.
It’s not clear how widely open source models are being deployed but it’s estimated there are now more than 2.5 million, many of which can be found on platforms such as Hugging Face. The one thing that is certain is that, in terms of capabilities, any gap that exists between proprietary and open source models is starting to narrow, especially as more time and effort is poured into advancing open source frontier models that include DeepSeek, Qwen, Llama 3 and Mistral Large 3.
In the meantime, the lack of clarity surrounding the definition of open source models should not discourage adoption, says Dolan. The role of open source has not diminished, he adds. In fact, as cost becomes a larger concern, more organizations will rely on open source models to automate certain tasks more cost effectively. Already, all the major cloud service providers make open source AI models available as an alternative to proprietary AI models. Right now, however, proprietary AI models are all too often just the default option for many organizations that have not spent much time researching other options.
The real issue is simply establishing a level of transparency that makes it clear that from a licensing and terminology perspective not all open models are created equal.
