DeepSeek says both models are more efficient and performant than DeepSeek V3.2 due to architectural improvements, and have almost “closed the gap” with current leading models, both open and closed, on reasoning benchmarks.