Cheap Heads, Not Bigger Bases: Training Beats Scaling for On-Device Function-Call Selection

On-device assistants must pick the right tool from a candidate set in a single forward pass, on a small model, under a latency budget. For this function-call selection task, the effective lever is to train a small base rather than scale it. We add a contrastive "selector head" (a DoRA adapter of about 0.14% of the parameters) to a frozen Qwen3-0.6B. This raises BFCL selection accuracy from 80.6 to 90.3, matching a vanilla model seven times its size. Training helps most on the smallest base and tapers as the base grows. Because the base stays at 0.6B, the efficiency reaches real hardware. On an Apple M1 the trained model runs about five times faster and smaller than the vanilla 4B it matches, and the whole study costs roughly 10¹⁷ FLOPs. The same recipe is competitive with frontier APIs on discrimination and far more reliable at knowing when not to call a tool, beating every frontier model there by more than 23 points. It also transfers to a real application: a voice-driven digital-audio-workstation (DAW) assistant and the REAPER scripting API. The BFCL pipeline, split, and adapters are released for reproduction.