DeepSeek open source upgrades basic components: simultaneous release of computing, communications an

Editor's note:

The underlying AI chip tools from the DeepSeek team are open source.

On the morning of September 30, DeepSeek announced open-source infrastructure components for the Huawei Lifting Platform, covering TileLang high-level language compilation tools, high-performance computing libraries and distributed communications libraries. According to DeepSeek, these components are one-to-one matches with the previous open-source components for the Invidia platform.

The release includes DeepGEMM, DeepEP, TileKernels, FlashMLA and DeepSelect, respectively, to perform matrix operations, cross-device communications, conventional vector computation and visitor storage, scarcity attention and data filtering, all of which are key components of the model's operation.

Huawei has also released some of the offline reasoning performance data for DeepSeek-V4.1-Flash, which covers model deployments, large-scale training, KV Cache pools and Agentic RL, among other practices in the CANN community.

Link to the open source project: