On April 24, the DeepSeek V4 series of large-scale models was officially released and simultaneously open-sourced, featuring two versions: V4-Pro and V4-Flash. This series of models has achieved key breakthroughs in dimensions such as ultra-long context processing, agent capabilities, and reasoning performance, leading in multiple metrics among open-source models and matching the overall performance of international top-tier closed-source large-scale models. The V4-Pro is positioned for high-end complex tasks, with a total of 1.6 trillion parameters and approximately 49 billion activated parameters. It has reached the optimal level among open-source models in the Agentic Coding agent programming evaluation, surpassing all publicly evaluated open-source models in mathematics, STEM disciplines, and competitive coding. The V4-Flash is designed for high-concurrency lightweight scenarios, with a total of 284 billion parameters and 13 billion activated parameters, offering faster response speeds and lower invocation costs. The new versions are equipped with a brand-new attention mechanism, reducing single-token reasoning computation to 27% of the previous generation through a hybrid compression coefficient mechanism, and decreasing KV Cache usage to 10%, supporting ultra-long contexts of up to 1 million tokens. Several local computing power and technology companies, including Baidu, Cambricon, Moore Threads, and Huawei Ascend, completed first-day adaptation and integration. Huawei Ascend's full range of super node products reduced reasoning overhead through optimized fused kernels, multi-stream parallelism, and quantization algorithms. Securities research reports indicate that the accelerated adaptation of domestic large-scale models to local computing power will drive rapid growth in the computing power chip industry chain, with domestic computing power chip shipments expected to double by 2026.