DeepSeek V4 architecture upgrade improves efficiency

DeepSeek V4 has achieved a significant cost reduction in long context windows through architectural upgrades, with both its Pro and Flash versions supporting ultra-long context windows of up to 1 million tokens. This efficiency leap is achieved through three key architectural innovations: a hybrid attention mechanism, training stability, and a main training optimizer.