A compact, efficient model that punches above its weight for its size. With only 800 million active parameters drawn from a 3 billion parameter architecture, it uses a mixture-of-experts style approach to stay lean at inference time. Handles a generous 128k token context window, which is unusually large for a model this small.