Can smaller language models effectively handle tool calling tasks?

There's been a lot of discussion lately about whether you really need massive models to handle complex tasks like tool calling. Tool calling—where a language model can request external functions or APIs to accomplish goals—has traditionally been the domain of large, well-resourced models. But what if you could compress that capability into something much smaller and more efficient?

I'm curious whether anyone here has experimented with smaller models (in the 20-30M parameter range) for practical applications. The appeal is obvious: faster inference, lower computational overhead, easier deployment on edge devices. But I wonder about the real-world tradeoffs. Does accuracy suffer significantly? Are there specific types of tool-calling tasks where smaller models excel, and others where they struggle?

For teams building applications that need tool integration—whether that's API calls, database queries, or external service connections—what's your experience been? Are you still defaulting to the largest available models out of habit, or have you found good results with more modest architectures?

I'm also interested in the technical side: what approaches actually work for distilling complex capabilities into smaller models? Is it purely about training data curation, or are there architectural tricks that help preserve functionality at smaller scales?