DeepSeek Previews New AI Model for Huawei Ascend Chips

Chinese AI lab DeepSeek has previewed a new version of its large language model designed to run natively on Huawei Ascend chips. The move directly counters US sanctions that restrict NVIDIA GPU exports.

Why This Is a Breakthrough

Most advanced AI models depend on NVIDIA’s CUDA platform. Huawei’s Ascend series (910B, 910C) uses a different software stack called CANN. DeepSeek successfully ported its model to this architecture without major performance loss.

Early previews show the Huawei Ascend AI inference solution achieves:

  • 85-90% of NVIDIA A100 throughput
  • 40% lower cost per token
  • Full 128K context window support

Sanctions Accelerate Domestic Innovation

With the US blocking NVIDIA’s H100 and A100 chips, Huawei’s Ascend became China’s best domestic alternative. However, software immaturity kept developers away. DeepSeek’s preview proves that sophisticated LLMs can now run entirely on homegrown hardware.

Technical Edge

DeepSeek implemented custom kernel optimizations for Ascend’s matrix units and mixed-precision training via Huawei’s MindSpore framework. The result: stable performance on code generation, reasoning, and math tasks.

What’s Next

A full release is expected by Q3 2026. DeepSeek also plans an open-source toolkit to help other developers migrate models to Ascend. This could accelerate China’s AI self-sufficiency.

Bottom line: DeepSeek turned a sanctions obstacle into a strategic advantage. The era of domestic AI hardware is arriving faster than expected.

Scroll to Top