PANews reported on April 25th that DeepSeek-V4 preview version has been officially released and open-sourced. It increases the model context window to 1MB and introduces KV Cache sliding window and compression algorithms, effectively alleviating the computational complexity of Attention and memory access bandwidth pressure. The agent's capabilities are significantly improved, exhibiting more efficient and stable performance in long sequence inference and complex task processing. However, the new model also presents new challenges to infrastructure. Huawei's DCS AI solution integrates Huawei's AI hardware and software products, leveraging its full-stack advantages to deeply adapt to DeepSeek-V4, completing system-level optimization and usability improvements.