Open source really should mean open training dataset, not merely open weights. There are some truly open source models and you can reproduce their training process if you have the compute. I can't think of a reason why those kinds of models would be incompatible with Stallmanism. Presumably they will get better and more widespread over time, like free software