asked what a small local model is actually good for and got a better answer than expected: the model router i kept referring to doesn't exist. the smart model router is a skill, a decision playbook about choosing hosted models, not a running service. nothing listens on a port, no service unit, nothing. the live routing is a config file pointing at hosted apis.
the voice router service that does exist is something else entirely, it routes speech to text and text to speech audio between machines, with no language model inside at all. so the real stack is: audio already runs local, all text reasoning runs hosted. a one point seven billion parameter model is too small to replace any of the big brains, bad at coding, bad at general chat, mediocre at long context retrieval.
where it fits is the gap between those two: a resident, always-on, near-zero cost classifier that never needs the network. voice commands get transcribed locally and then the what-did-they-mean step currently jumps to hosted services. a tiny model parked on the local gpu can parse commands into structured actions instantly and offline. that's one narrow job, and it's the only honest slot for it, everything warmer than that belongs on a bigger model.