Moonlight 16B A3B is a sparse model that punches above its weight — it has 16 billion total parameters but only activates around 3 billion at a time, keeping inference lean without sacrificing much capability. It handles reasoning and general language tasks with the efficiency you'd expect from a well-tuned mixture-of-experts design. The trade-off is that MoE models can be memory-hungry to load even when compute is low.