DeepSeek and Huawei ship open-source tools for Ascend 950

The release arrived with a concrete technical marker: TileLang added Huawei Ascend 950 as an official backend on September 30, 2026, joining Nvidia CUDA, AMD ROCm and Apple Metal. Beyond the language itself, DeepSeek shipped libraries that cover matrix multiplication, inter-NPU communication for Mixture of Experts models, and attention processing used in DeepSeek-V4.1 inference, plus a set of dozens of high-speed kernels that automatically detect whether an Nvidia GPU or Huawei NPU is running.

Huawei fully supported the development effort, and the two companies also worked on optimizing a "supernode" configuration carrying 128 Ascend 950 chips, tuning both compute and chip-to-chip communication. DeepSeek's stated goal is a high-level language that is general-purpose, easy to program and still able to extract hardware performance for AI accelerators.

The competitive read is that this is one of the clearest efforts yet to build a software layer that lets AI workloads run on Huawei silicon without rewriting code written for Nvidia's ecosystem, though the practical test will be whether real-world workloads match the reported numbers — FlashMLA reaching 95% of theoretical performance on some operations and DeepSelect's 2x to 20x speedup over torch.topk. For teams currently tied to CUDA, the toolchain may offer a lower-friction migration path, but only if the performance claims hold outside DeepSeek's own benchmarks.