You are not alone. It blew my mind at how good it is per billion parameters. As an example, I can’t think of another model that will give you working code at 4B or less. I havent tried it on agentic tasks but that would be interesting