Yes I think very constrained task: known data universe, well known language etc should be the best possible place for small language models to play
Another point, but I feel like, we can see some really good fine tuned models out of this model, the community feels excited about 1-bit LLM architecture. We are gonna see some good innovation within this space in the upcoming future.