DeepSeek delays V4 for Ascend
- DeepSeek released a preview of its V4 artificial intelligence model on April 24 after delaying the broader rollout while tuning the system to run on Huawei’s Ascend chips. - The preview comes in two versions, V4-Pro and V4-Flash; Reuters said V4-Pro has 1.6 trillion total parameters, 49 billion active parameters, and a one-million-token context window. - The launch ties DeepSeek’s flagship model more closely to Huawei hardware as China pushes a more self-sufficient artificial intelligence stack under tighter U.S. chip restrictions. (reuters.com)
DeepSeek did not scrap V4. It released a preview on April 24 after holding back a fuller rollout while tuning the model for Huawei’s Ascend chips. (reuters.com) (bloomberg.com) The Chinese startup said V4 comes in two versions, DeepSeek-V4-Pro and DeepSeek-V4-Flash, and kept the release open-source so developers can download and modify the model. (cnbc.com) (reuters.com) Reuters reported that V4-Pro has 1.6 trillion total parameters with 49 billion active parameters, while V4-Flash has 284 billion total parameters with 13 billion active parameters. Both versions support a one-million-token context window. (reuters.com) (tech.yahoo.com) A context window is the amount of text a model can keep in view at once, like the size of a desk before papers start falling off. DeepSeek is pitching that larger window for agent tasks, coding, knowledge retrieval, and long documents. (cnbc.com) (reuters.com) The hardware angle is the reason this story moved. Bloomberg reported that DeepSeek spent months reworking its software stack for Huawei-specific tuning instead of settling for basic compatibility. (bloomberg.com) Huawei said on April 24 that its Ascend supernode, built on Ascend 950 artificial intelligence chips, would fully support DeepSeek’s V4 versions after the preview launch. Reuters said that marks a sharper turn from DeepSeek’s earlier reliance on Nvidia chips. (reuters.com 1) (reuters.com 2) That shift sits inside a larger supply-chain squeeze. U.S. export controls have limited Chinese access to the most advanced foreign artificial intelligence chips, pushing Chinese model makers toward domestic hardware and software tools. (reuters.com) (bloomberg.com) Analysts quoted by CNBC said V4 appears aimed less at shocking markets than R1 did in early 2025 and more at lowering inference costs and improving agent performance. Morningstar’s Ivan Su told CNBC traders had already priced in the idea that Chinese models can be competitive and cheaper. (cnbc.com) So the cleanest way to read the delay is this: DeepSeek slowed the public rollout, shipped a preview anyway, and used the extra time to bind its next flagship model more tightly to Huawei’s chip stack. (bloomberg.com) (reuters.com)
