Moonlight 16B A3B is a sparse mixture-of-experts model that activates only 3 billion parameters per forward pass despite having 16 billion total — meaning it punches above its active-parameter weight in reasoning tasks while staying computationally lean. It handles instruction-following with a focused, efficient style, trading breadth for speed. Like a specialist who knows when to delegate, it routes computation selectively rather than brute-forcing every token.