DeepSeek V4 launches with 1.6T parameters on Huawei Ascend chips
1.6 trillion parameters, 1M token context, day-zero Ascend support.
DeepSeek has officially launched DeepSeek V4, a massive Mixture of Experts (MoE) model with 1.6 trillion parameters and a 1 million token context window. In a strategic move, the model achieves day-zero support on Huawei's Ascend 950DT chips, thanks to close collaboration between the two companies. This integration ensures optimal inference performance and reduces dependency on other hardware vendors, signaling a significant step toward self-sufficiency in China's AI infrastructure.
Developers and enterprises can now deploy DeepSeek V4 on Huawei Ascend clusters immediately, leveraging the model's vast capacity for complex reasoning, long-document analysis, and multi-turn conversations. The partnership aims to offer a competitive alternative to NVIDIA-powered solutions, with DeepSeek claiming tailored optimizations for the Ascend architecture. This launch could accelerate adoption of domestic chips for large-scale AI workloads.
- 1.6 trillion parameters with MoE architecture for efficient scaling.
- 1 million token context window enabling extended document processing.
- Day-zero optimization on Huawei Ascend 950DT, reducing reliance on other chipmakers.
Why It Matters
Reduces dependence on NVIDIA hardware, empowering Chinese AI deployments with domestic chip support.