DeepSeek open-sources a Huawei Ascend kernel stack to rival CUDA
Free download: TileLang, DeepGEMM, FlashMLA, DeepEP and more, tuned for 128-chip Ascend 950 supernodes — the open software layer non-Nvidia silicon lacked.

Copy markdown
DeepSeek dumps its Ascend kernel toolchain
Announced Sept 30 via DeepSeek's WeChat: TileLang (a Python-style, CUDA-like kernel DSL) plus DeepGEMM for matrix multiply, FlashMLA for memory, TileKernel for compute, DeepSelect, and DeepEP for inter-chip comms, with benchmarking tools — all free to download.
TileLang isn't Huawei-only
TileLang is an open, TVM-backed Python DSL for writing high-performance kernels across accelerators — a Triton-like path, not vendor lock-in. The new part is a production-grade Ascend backend, so the same kernel style now reaches Huawei silicon too.
Built for 128-chip supernodes
Everything is tuned for Ascend 950 in 'supernode' clusters of up to 128 accelerators, optimizing compute and inter-chip communication together. It generalizes DeepSeek's April 2026 work porting V4 to Huawei hardware into tooling anyone can pick up.
Why it lands on your compute bill
A real open software layer is the piece non-Nvidia chips have lacked; with it, Ascend becomes a more viable place to train and serve, adding downstream pressure on GPU cost and supply. It's a higher-level portability layer, not a CUDA killer yet — but the moat just narrowed.