Moming Duan

The current status quo is that some people apply licenses like CC-BY-NC to their models to prohibit commercialization, which I believe creates more problems for the AI ecosystem. In my view, it’s OK for people to choose not to use an open source license when publishing their models, but there needs to be a clear and visible way to distinguish open source models from non–open source ones.

MG-BY-OS is a copyleft license, and I envision it serving a similar role for models as GPL-3.0 does for software. But its adoption depends on the voice and consensus of the ML community.

I see, but I’d like to clarify that we have only submitted MG0, MG-BY, and MG-BY-OS for open source approval by OSI. I acknowledge that other ModelGo variants do not comply with the OSD and FSF’s definition of free software. However, I believe it’s still important to provide users with options—such as RAI, NC, or ND—for publishing their models, rather than pushing them toward less suitable alternatives like proprietary licenses (e.g., LLaMA 2 or Gemma), which often introduce more legal ambiguity.

I am not arguing that every model must be open source, but it should be easy to clearly distinguish open source models from those that are not.

I appreciate this article and have also observed the widespread indifference toward model licensing on this platform. I’ve been working with lawyers to draft a new set of model-specific license (similar with model style of CCs) called ModelGo for general model publishing. We’ve submitted several variants of ModelGo for review by OSI and hope it can help establish a more standardized way to model sharing.

Note: The current version of ModelGo Licenses on our website is not the latest one submitted to OSI, as we are still addressing their feedback.