DeepSeek V4 will be released in the coming weeks and may run entirely on Huawei chips.

BlockBeats message. April 4, according to the latest report by US tech media The Information, China’s AI startup DeepSeek is about to launch its next-generation flagship model V4, which will run entirely on Huawei’s in-house developed chips. It is seen as an important milestone for China’s push toward semiconductor self-reliance.

V4 is expected to be released within the next few weeks and will run entirely on Huawei chips. For this purpose, DeepSeek has worked with Huawei and chip designer Cambricon for several months, rewriting parts of the model’s underlying code to achieve compatibility with domestic chips.

Notably, this time DeepSeek did not grant NVIDIA early testing access for V4, and instead only invited domestic chip companies to participate in early optimization—breaking industry convention.

Chinese tech giants such as Alibaba, ByteDance, and Tencent have already pre-purchased Huawei’s latest Ascend 950PR chips, with total order volume reaching hundreds of thousands of units. They plan to deploy V4 via cloud services and integrate it into their own AI applications. Surging demand has driven the chip’s price up by about 20%.

V4 uses a Mixture of Experts (MoE) architecture, with a total parameter count of about 1 trillion. For each inference, it activates about 37 billion parameters, supporting multimodal inputs of text, images, and code while maintaining low latency.

Last year, DeepSeek released the low-cost models V3 and R1, which triggered a global selloff in technology stocks and led the market to question whether US AI companies need to spend tens of billions of dollars to buy compute power. V4 has therefore attracted significant international attention.