You can download the AI model. How open is it really?

When someone describes an AI model as open, ask what they mean. Can you download the parameters? Inspect the code? Understand the training process? Modify the system? Use it in your intended setting? These are different questions, and one answer does not automatically settle the others.

The Open Source Initiative’s Open Source AI Definition 1.0 addresses freedoms to use, study, modify and share a system. It also describes requirements for parameters, code and data information. Under that definition, available weights by themselves are not the whole picture. This is the OSI’s definition, not a claim that all organisations use the term identically. Read the definition

For a fictional team considering a model, the useful exercise is a small inventory. List the components you can access and the terms attached to each. Then compare that list with the job: experimentation, internal deployment or redistribution. Marketing shorthand should not substitute for reading the applicable licence and documentation.

Our view: describing the actual access is more useful than awarding an “open” badge. A clear statement about weights, code, documentation and restrictions helps readers compare systems on their own terms. Openness is an important conversation. It gets better when the vocabulary is specific.