At Huawei Connect 2025, open-source AI development emerged as a significant focus, with Huawei revealing comprehensive plans to make its AI software stack publicly accessible by the end of the year. The company provided a clear timeline and detailed technical specifics to assure developers and address previous friction encountered with Ascend infrastructure.
Eric Xu, Huawei’s Deputy Chairman, acknowledged past challenges developers faced, particularly with the Ascend 910B and 910C chips. Customer feedback has played a critical role in shaping the open-source strategy intended to tackle these issues by enhancing community involvement and transparency.
A key component of this initiative is CANN (Compute Architecture for Neural Networks), Huawei’s core toolkit interfacing between AI frameworks and Ascend processors. Huawei plans to open interface for the compiler and virtual instruction set by the end of December 2025, ensuring developers gain insights into the compilation processes for performance optimization. While maintaining some proprietary elements, this move aims to balance transparency with performance.
In addition, Huawei committed to fully open-source their Mind series application enablement kits and toolchains, allowing developers to modify and expand the tools they regularly use. This commitment signifies a substantial shift toward community-driven development and ecosystem evolution.
Huawei also plans to open-source its openPangu foundation models, joining other companies in the open-source foundation model domain. These models, crucial in AI application development, still lack specific details like parameter counts and licensing arrangements, which are anticipated upon their release.
Operating system compatibility was another crucial aspect addressed during the event. Huawei announced the open-sourcing of its UB OS Component, enabling integration with existing operating systems like openEuler. This flexible approach allows organizations to embed Huawei’s components into their systems, reducing deployment friction and fostering easier use of its infrastructure.
Compatibility with existing AI frameworks remains critical for developer adoption. Huawei emphasized its support for frameworks like PyTorch and vLLM, aiming to allow developers to continue using familiar tools seamlessly with Ascend hardware.
Huawei’s ambitious timeline of a December 31, 2025 release, covers CANN, the Mind series, and openPangu models, suggesting significant work is already underway. For the initiative’s success, Huawei will need to ensure comprehensive documentation and sustained community support post-release to build an active contributing community.
Several specifics, such as licensing terms and governance structures, remain undefined but are vital for determining the extent of community involvement and commercial application potential. Developers and organizations now have a window before the December release to evaluate whether Huawei’s open-source AI platform aligns with their needs, with mid-2026 seen as a possible timeframe for discernible community adoption patterns to emerge.