Huawei is doing great with domestic development and in-house Ascend AI chips, while DeepSeek just dropped its powerful new V4 model, a huge 1.6 trillion parameter AI with a Mixture of Experts design and support for 1 million token context.
What’s exciting for Huawei users? It worked really well on Ascend hardware right from the first day. On launch day, the Huawei Ascend 950DT performed really well.
Huawei and DeepSeek worked closely together, so the CANN software was ready with fused operators and smart use of the dual-core design. This helped the model run smoothly without the usual early problems.
Thanks to this teamwork, DeepSeek official API started serving V4 on Huawei Ascend from Day 0. That kind of instant support is rare for such a big open-source model.
The model uses FP4 precision that matches Ascend’s strengths perfectly. Because of this, enterprises and cloud providers could start using it quickly on Ascend SuperNode clusters.
While some other hardware struggled with slow speeds or missing optimizations, Huawei gave practical, usable performance from the beginning.
SemiAnalysis report clearly shows how Huawei’s hardware-software co-design shines for modern MoE models like DeepSeek V4.
This is more than one successful launch. It proves Ascend chips are ready for cutting-edge AI models at scale.
For companies in China and beyond, it means faster, more independent, and more efficient AI deployment.
DeepSeek V4 on Ascend 950DT is a great example of what’s possible when hardware and software teams work hand in hand.
(Image Credits: Huawei Blog)
The post DeepSeek V4 and Huawei Ascend: A powerful day 0 partnership that shows the future of AI appeared first on Huawei Central.
