The useful ones are still provided by big companies
You're falling for the fallacy that larger parameter size = more useful. That's just a marketing gimmick by the AI industry to justify larger investments.
These "flagship" models with hundreds of billions of parameters and giant context windows have not achieved proportional gains in benchmarks.
You know what does significantly improve accuracy, even on smaller-sized LLMs? Retrieval-Augmented Generation. There's literally no reason to use giant, resource-intensive models. You can just give a smaller one access to databases and libraries with all the information it needs to report on.
A 24 billion parameter model is easily self-hostable on consumer hardware, and can be quantized to further reduce hardware requirements for a marginal loss in accuracy. At 8-bit quantization, that only requires 24GB of RAM + overhead. A lot by some standards, but by no means unachievable for a hobbyist. For a medium-sized business, that's downright negligible. And you can easily expand your context window using a swapfile.
If you don't have that much RAM, a 12 or 14 billion parameter model, even at 8-bit quantization, is fine if you do RAG and use swap to expand the context window. You can run that even if you only have 16GB RAM total on your hardware.
the rest of us can't afford the hardware to train them.
Most of the models on huggingface are pre-trained and list their datasets. You can fine-tune and align them yourself, which uses far less resources than pre-training does. You can even use LoRA to further reduce resource needs.
Don't fall for the lie that commercial LLM APIs are the only viable option.