Open models look a lot like open source for AI: you can download them, run them on your own hardware, modify them, and build them into your product without asking a vendor for permission. Yet by strict definitions they are not fully open source, because training data and training code usually stay closed. So are they the open source of AI, or not yet?
For several years I have been using open models on devices: phones, desktops and the web, without the cloud. The weights were almost never the problem. What got in the way was everything around them: the right format didn’t exist, there was no way to run the model on the platform I needed, or no way to fine-tune and convert it for my task. These gaps were filled not by the model authors but by the community, through runtimes, frameworks, converters and evaluation tools built by ordinary contributors.
In this talk, I’ll look at which open source freedoms open models already give us and which are still missing. Using my own experience and my projects, I’ll show that open source in AI is more than weights, and where in this stack you can contribute today.
