Observed

Holds for capable models in non-streaming mode: native GBNF grammar enforcement since v0.5; Qwen3-14B reaches F1=0.971 tool selection (source-reviewed corpus benchmark). Materially qualified: above ~5-6 active tools qwen3-coder switches from JSON to XML tool-call format and some models stop attempting tool calls entirely — silent, no error signal; tool-calling support is model-dependent on the MLX backend (March 2026 Apple Silicon switch).