DeepSeek launches V4 on Huawei chips

DeepSeek has released a preview of V4, a new open-source large language model built to work with Huawei’s Ascend chips. (reuters.com) The Hangzhou startup published two versions on April 24: DeepSeek-V4-Pro and DeepSeek-V4-Flash. DeepSeek said V4-Pro outperforms other open-source models on some world-knowledge tests and trails only Google’s closed-source Gemini-Pro-3.1 on that benchmark. (reuters.com) (scmp.com) V4-Pro has 1.6 trillion parameters, while V4-Flash has 284 billion. Both models can handle up to 1 million tokens of context, up from 128,000 in DeepSeek’s previous flagship model. (scmp.com) (huggingface.co) A context window is the amount of text a model can keep in view at once, like the size of its working memory. DeepSeek is pitching that larger window for agent tasks, coding tools and long documents rather than just chatbot replies. (cnbc.com) (api-docs.deepseek.com) The hardware angle is the bigger shift. Reuters reported Huawei said Ascend chips were used in some of V4’s training, and Huawei later said its Ascend chips and supernode systems offered “full support” for running V4 in inference. (reuters.com) (scmp.com) That matters because Nvidia still dominates the global market for training and serving advanced artificial intelligence models. DeepSeek’s earlier systems were associated with Nvidia hardware, so V4 marks a more public alignment with China’s main domestic alternative. (reuters.com) (channelnewsasia.com) The release lands as Washington tightens export controls and the White House accuses China of large-scale theft of United States artificial intelligence know-how. Nvidia chief executive Jensen Huang warned this month that losing Chinese developers to Huawei would be “a horrible outcome” for the United States. (reuters.com) DeepSeek is also using price as a weapon. Its API page lists V4-Flash at $0.14 per 1 million uncached input tokens and $0.28 per 1 million output tokens, while V4-Pro is temporarily discounted 75% through May 5. (api-docs.deepseek.com) Early reaction has been split. CNBC cited Counterpoint analysts saying V4 shows strong agent capability at lower cost, while Reuters quoted Hugging Face engineer Lewis Tunstall saying it is strong on very long text tasks but lacks image and video support. (cnbc.com) (reuters.com) DeepSeek’s last market-shaking release was R1 in January 2025, which the company said was built in two months for under $6 million using lower-capacity Nvidia chips. Analysts told CNBC V4 is unlikely to hit markets the same way, but it gives China a new test of whether top-tier models can move onto Chinese hardware at scale. (cnbc.com) (reuters.com)