DeepSeek released V4 AI models, claiming they nearly match top US rivals. The V4-Pro trails only Google's Gemini 3.1-Pro in world knowledge and is "marginally short" of OpenAI's GPT-5.4, while beating Anthropic's Sonnet 4.5. The models were tested on Huawei's Ascend NPUs, signalling a shift in China's AI stack.