How can unslothai improve a model (decrease the ram, the context window....)

[vricosti]

Hello,
I’d like to understand how a company like UnslothAI manages to reduce the amount of RAM used by an LLM model.
If it were that easy, why wouldn’t the original LLM designers do it directly?

It feels a bit like a company doing the same thing with cars by unlocking the engine — yes, it’s possible, but there’s always a price to pay in that case. I’d like to understand what the downsides are when doing this with an LLM.
Otherwise, it just sounds too good to be true.