We're exploring sub-3B models for on-device inference (latency and privacy reasons). The quality gap vs. cloud models is obvious, but for narrow tasks it might be fine. Anyone shipping small models on edge hardware? What model sizes, quantization levels, and frameworks proved reliable?