Welcome to The New Stack Agents, a Weekly Livestream on AI
Open source AI is a fast-growing community and the PyTorch Foundation, not even three years old yet, is emerging as a hub for those creators and maintainers. It now sponsors vLLM, an inference and serving engine for large language models (LLMs), and DeepSpeed, a suite of deep-learning optimization tools, as well as PyTorch, an open source machine learning (ML) framework that originated at Meta in 2017.
The PyTorch Foundation’s Executive Director, Matt White, spoke with The New Stack Agents host Frederic Lardinois about the growth of the PyTorch ecosystem and its introduction of OpenMDW, a new open source license for AI models. The livestream was recorded in Denver at Open Source Summit North America.
White, who is also general manager of AI at the Linux Foundation, said he has been “uniquely blessed” to be involved with the joining of two worlds — AI and open source.
“There are a lot of open source folks that are coming into this, and they don’t really understand the nuances of AI,” he said. “And then there’s AI folks that don’t understand licensing, they don’t understand open source, right? And so the converging of these two worlds is obviously driving out a lot of things, right? We see standards starting to emerge. Folks are looking at specifications that’ll be more widely adopted for particular types of general use cases.
“This is the really neat part of this — that not only we’re moving fast, but we’re also coming together, and we’re figuring out things together.”
What’s the OpenMDW License?
One thing the two sides are figuring out together, via the PyTorch Foundation, is licensing for open source AI. In May, the foundation announced OpenMDW, short for Open Model, Data and Weights license.
“We quietly released it a few weeks ago, and not that quiet,” White told the TNS Agents audience. He added, “It’s gone through a lot of scrutiny, and now it’s just the ‘first mover’ problem. Like, let’s see who wants to believe in this.”
The OpenMDW license came about, White said, to answer an industry call for an AI-specific and model-specific license. “There’s sort of two camps right now where we call like, the open weights models, which are using community license with restrictions in them.” The MIT license is an example, he said.
“Then we have the permissive license and copy-left licenses, which are the [Open Source Initiative]-approved licenses, which we would call open source licenses.” The Apache license would be an example.
But, White added, “most labs and most organizations recognize that those aren’t the most appropriate licenses, because they don’t talk about the artifacts, the disclaimers that are involved with AI models.”
To come up with something more suited to open source AI models, White said, they sought feedback from a broad spectrum of users. “We went to the community and said, ‘Hey, we think we need something here.’ Work with the community, work with a lot of member organizations and academia to write something that was going to be very succinct and brief.
The result, he said, is “like a one-pager,” guidelines that cover “all of the potential trademark and patent and database rights and these sort of things, to make sure that it’s well understood that when you’re getting the model, you can use it for any purpose. You can study, modify, redistribute, quantize.”
Check out the full episode of TNS Agents for more on OpenMDW, the open source Agent2Agent and Model Context Protocol (MCP), and what White sees ahead for the fast-moving AI space.
