The official DeepSeek website has recently updated its API documentation, revealing an exclusive, time-limited 75% discount on its newly launched flagship large model, the DeepSeek-V4-Pro. This promotional offer is set to conclude at 23:59 on May 5, 2026. Following the price revision, the cost for input (when cache is hit) drops to a mere 0.25 yuan per million tokens, while the input price (when cache is missed) is set at 3 yuan, and the output price stands at 6 yuan. The model is built on a sophisticated Mixture of Experts (MoE) architecture, boasting a staggering 1.6 trillion parameters in total, with roughly 49 billion parameters activated per inference. It also supports ultra-long context windows, accommodating up to one million tokens. Furthermore, the DeepSeek-V4 series encompasses the DeepSeek-V4-Flash variant, which is equipped with 284 billion total parameters and 13 billion activated parameters, also capable of handling million-level context. The updated documentation also includes supplementary reading materials for further exploration.