← Back to Blog
EN2026-10-14

DeepSeek and Huawei Open Source More of the Ascend AI Stack

DeepSeek's new Ascend-focused repositories show how AI infrastructure is moving from model releases into kernels, communication libraries, and hardware-specific developer tooling.

By NeoAI
AIOpen SourceDeveloper ToolsAI InfrastructureDeepSeek

The most interesting AI release this week is not another chatbot. It is lower in the stack: kernels, communication libraries, and compiler support for Huawei's Ascend accelerators.

On September 30, 2026, DeepSeek published DeepGEMM-Ascend, an Ascend implementation of its matrix multiplication kernel library, with initial support for Ascend 950 devices.1 Around the same release wave, DeepSeek also published DeepEP-Ascend, a communication library for machine learning training and inference on Huawei Ascend NPUs.2 Reporting from Tom's Hardware, citing Reuters, described the broader package as open-source programming tools developed with Huawei support, including compute libraries, chip-to-chip communication components, and Ascend support for TileLang.3

That matters because large AI systems are not only limited by model architecture. They are limited by how efficiently each accelerator performs common operations and how quickly data moves between accelerators. Matrix multiplication is the obvious compute primitive. Mixture-of-experts models add another pressure point: routing tokens to different experts and combining the results without leaving expensive processors idle.

DeepGEMM-Ascend is aimed at the first problem. Its README says the library keeps the same package name and API shape as DeepSeek's existing DeepGEMM library, while supporting BF16, FP8, FP4 GEMM, MQA logits, and MegaMoE operators on Ascend.1 The project describes a lightweight abstraction over Ascend matrix multiply-add primitives, hiding details such as fractal layouts, alignment constraints, address calculations, and low-level parameter handling.1

DeepEP-Ascend is aimed at the second problem. The repository describes it as a high-performance communication library for training and inference, providing expert-parallel all-to-all operations for MoE dispatch and combine, with FP8 dispatch and deferred epilogues.2 It also lists work-in-progress primitives for pipeline parallelism, context and data parallelism collectives, and remote memory access.2

TileLang is the third piece. The TileLang Ascend 950 backend guide says the Ascend dialect reuses TileLang's shared frontend, with Ascend-specific lowering, scheduling, synchronization, and code generation for Cube and Vector execution.4 In practical terms, that gives developers a higher-level way to write kernels for Ascend hardware without writing every operation directly against the lowest-level platform interfaces.

The important nuance is that this is not a general-purpose replacement for CUDA in the way a headline might imply. The tooling is specific. DeepGEMM-Ascend's requirements include Huawei Ascend NPUs, the CANN 9.20 toolkit, torch_npu, Python 3.10 or higher, and C++20 build support.1 DeepEP-Ascend is similarly tied to the Ascend software stack, and its README notes that some published performance results were collected on a proof-of-concept hardware kit, not an unreleased commercial baseline.2

That limitation does not make the release unimportant. It makes it more precise. DeepSeek and Huawei are not just releasing models; they are filling in the infrastructure needed to run large models on non-Nvidia hardware. For developers, the message is that AI portability is becoming a full-stack problem. Model weights are only one layer. The real lock-in lives in compilers, kernels, collective communication, profiling tools, framework integrations, and the operational knowledge required to make all of them fast.

The open-source angle is also practical. These repositories let researchers and infrastructure engineers inspect how Ascend-specific optimization is being approached, even if most teams do not have Ascend 950 hardware on their desks. DeepGEMM-Ascend exposes the compute path. DeepEP-Ascend exposes the communication path. TileLang exposes a higher-level kernel authoring path.

For the broader AI ecosystem, this is a useful signal. The next phase of competition is not only who has the strongest model. It is who can make the developer stack around alternative accelerators credible enough that engineers can actually build, debug, and optimize serious workloads on them. Open source does not solve that by itself, but it gives the ecosystem something concrete to test instead of another promise slide.

Sources

Footnotes

  1. DeepSeek AI, "DeepGEMM-Ascend", accessed October 14, 2026. ↩ ↩2 ↩3 ↩4
  2. DeepSeek AI, "DeepEP-Ascend", accessed October 14, 2026. ↩ ↩2 ↩3 ↩4
  3. Tom's Hardware, "DeepSeek and Huawei release open-source Ascend AI programming tools to reduce reliance on Nvidia CUDA ecosystem", October 1, 2026. ↩
  4. TileLang, "TileLang Ascend 950 Backend", accessed October 14, 2026. ↩
intelliBrain

AI-augmented software development. Based in Zürich, working globally.

© 2026 intelliBrain GmbH. All rights reserved.Imprint
BUILT WITH 🧠 + AI