You'd need millions of dollars to just compile and label datasets. The training itself requires a lot of resources and money, as does human reinforcement.

Open source models would need benefactors with deep pockets.