The definitive version of DeepSeek V4 is poised for release, with the launch date projected to be as early as tomorrow and certainly within the coming days. A select group of users has already been invited to participate in grayscale testing. This launch encompasses two distinct iterations: DeepSeek V4 Flash and DeepSeek V4 Pro. DeepSeek V4 Pro boasts an impressive 1.6T parameters in total, with 49B actively engaged parameters, and can support a context length extending up to 1M (one million) Tokens. It shines particularly in domains such as Agent functionalities, comprehensive world knowledge, and superior reasoning capabilities. On the other hand, DeepSeek V4 Flash, featuring a total of 284B parameters and 13B actively utilized parameters, is tailored for general conversational interactions and fundamental text-based tasks that necessitate swift responses. Furthermore, the DeepSeek team has revealed that API pricing strategies will undergo adjustments in tandem with the official debut of V4, introducing a peak-valley pricing system. Prices during peak demand periods will be set at double the standard rate, with the intention of incentivizing users to utilize services during off-peak hours through pricing differentials, thereby mitigating server strain during high-traffic periods and bolstering overall service reliability.