DeepSeek Partners with Huawei to Open Source 6 AI Development Tools for Ascend, Including TileLang

The released software includes "TileLang," a domain-specific language designed to assist in developing high-performance computational programs, and "DeepGEMM Ascend," a matrix computation library. DeepSeek explained that they aim to provide a programming environment that maximizes hardware performance with more concise syntax than CUDA (NVIDIA's parallel computing platform).

Playing a crucial role in development, TileLang is designed to allow developers to write small computational programs called "kernels"—which run on GPUs and NPUs (AI processors)—using relatively simple code. As of September 30, 2026, the Ascend 950 has also been officially supported as a backend.

Other tools have also been released to accelerate AI model training and inference:

• DeepGEMM Ascend: A library for high-speed matrix multiplication. It supports computations such as BF16, FP8, and FP4, and allows the use of existing DeepGEMM APIs.
• DeepEP Ascend: A communication library that accelerates data communication between NPUs used in large language model training and inference. It is intended for use in environments such as Mixture of Experts (MoE).
• TileKernels: A library comprising dozens of high-speed kernels implemented using TileLang.
• FlashMLA: A library for accelerating attention processing, which has reportedly reached 95% of theoretical performance on the Ascend 950.
• DeepSelect: A library for accelerating TopK processing. Used in DeepSeek's Sparse Attention and other methods, it claims to enable speedups from 2 to 20 times faster than "torch.topk" depending on the conditions.