You can pay $0 for those models because a company paid $lots to train them and then released them for free. Those models aren't going away now of course, but lets not pretend that being able to download the product of millions of dollars worth of training completely free of charge is sustainable for future developments. Especially when most of the companies releasing these open models are wildly unprofitable and will inevitably bankrupt themselves when investments dry up unless they change their trajectory.
Much could be said about open source libraries that companies release for free to use (kubernetes, react, firecracker, etc). It might be strategically make sense for them so in the meantime we’ll just reap the benefits.
All of these require maintenance, and mostly it's been a treadmill just applying updates to React codebases. Complex tools are brittle and often only makes sense at the original source.
You’re acting as if computing power isn’t going to get better. With time training the models will get faster.
Let me use CG rendering as an example. Back in the day only the big companies could afford to do photoreal 3D rendering because only they had access to the compute and even then it would take days to render a frame.
Eventually people could do these renders at home with consumer hardware but it still took forever to render.
Now we can render photoreal with path tracing at near realtime speeds.
If you could go back twenty years and show CG artists the Unreal Engine 5 and show them it’s all realtime they would lose their minds.
I see the same for A.I., now it’s only the big companies that can do it, then we will be able to do it at home but it will be slow and finally we will be able to train it at home for quick and cheap.
The flipside to that metaphor is that high-end CG productions never stopped growing in scope to fill bigger and better hardware - yes you can easily render CG from back in the day on a shoestring budget now, but rendering Avatar 2 a couple of years ago still required a cluster with tens of thousands of CPU cores. Unless there's a plateau in the amount of compute you can usefully pour into training a model, those with big money to spend are always going to be several steps ahead of what us mere mortals can do.
Llama, Mixtral, Stable diffusion and Flux are a lot of fun and free to run locally, you should try them out.