I have spent some time evaluating small language models (350M parameters) and they are significantly different from LLMs. It seems that using these small models is not at all like using an LLM. The models can create new possibilities, but they are much more difficult to use.
Basically small models are a pure cost and efficiency optimization over large language models. A large model can do everything a small one can, and better; the small model is less able to understand what is necessary.
I feel the best thing about small language models is that they can be used by nearly anyone with a computer. Large language models meanwhile seem to be pay-to-play: you have to either pay money to a company each month for access, or own a lot of expensive hardware to run one locally.
I have always admired how computing equalizes people and creates opportunities. Anyone with a computer could learn to program. With LLMs, we now have a pay-to-play system where only those who can afford the tool can do certain tasks. Even though small models are not a replacement for LLMs, they bring us back to a more equal playing surface.