They wrote a whole ass paper about SLMs that do specifically this - expert small language models with narrow expertise.

And then went for a massive (but private and secure) datacenter instead.