DeepSeek released V4, an open-source flagship model that supported a 1 million-token context window, used a memory-efficient attention design to reduce compute and memory for long contexts, and was optimized for inference on domestic Chinese chips such as Huawei's Ascend.