pub const MODEL_PRIMARY: ModelSpec;
Primary model: Qwen3-4B-Instruct, Q4_K_M (~2.5 GB), used when RAM ≥ ~6 GB.