On April 24, 2026, the long-awaited preview version of DeepSeek V4 was formally launched, accompanied by the open-sourcing of its model weights and the release of a comprehensive technical report. The V4 series unveiled this time comprises two distinct editions: V4-Pro and V4-Flash. Both editions boast the capability to process an ultra-long context of 1M (one million tokens). V4-Pro shines in its Agent functionalities, extensive world knowledge repository, and superior reasoning prowess, positioning it on par with leading closed-source models. On the other hand, V4-Flash is designed as a streamlined, cost-effective variant. It offers reasoning capabilities akin to V4-Pro but with a more compact parameter size, enabling swifter and more economical API services. Presently, V4 has undergone thorough adaptation and optimization in collaboration with mainstream Agent products and is seamlessly compatible with domestic computing platforms, including Huawei Ascend.