“In a one-million-token context setting, DeepSeek-V4-Pro requires 10% of the KV cache compared to DeepSeek-V3.2.”