China's DeepSeek releases preview of long-awaited V4 model as AI race intensifies

deepseek released a preview of its v4 large language model on friday, over a year after its r1 reasoning model shook global tech markets. the hangzhou-based startup claims v4 performs strongly against domestic rivals, especially in agent tasks, knowledge processing, and inference. it's open-source, like v3, and optimized for tools like claude code and openclaw. two versions are available: "pro" and "flash." deepseek's r1 model, released in january 2025, alarmed investors after the company revealed it cost under $6 million and took two months to train on lower-capacity nvidia chips. that raised questions about the u.s.'s ai lead and big tech's massive infrastructure spending. since then, deepseek has released several upgrades, but none matched r1's impact. competition in china's ai sector is heating up — alibaba and bytedance also released new models this year. on friday, shares of chinese ai players like minimax and zhipu each fell around 8%, while manycore tech dropped 9%. a key question: which chips trained v4? huawei confirmed its ascend ai cluster can support the model, but it's unclear how much deepseek used huawei's chips versus nvidia's, given u.s. export controls.