Shenzhen, China – At its flagship Huawei Connect 2025 conference last week, Huawei laid out an ambitious roadmap to open-source its full AI software stack by December 31, 2025, signaling a major shift in how the company wants developers to engage with its Ascend AI ecosystem.
The plan includes releasing critical software layers, application toolchains, and even foundation models under open access — a move Huawei says is aimed at lowering barriers for developers and addressing long-standing frustrations with its infrastructure.
Admitting Friction, Promising Openness
Eric Xu, Huawei’s Deputy Chairman and Rotating Chairman, opened with an unusually candid acknowledgment: developers working with Ascend chips and tools have faced steep hurdles. He noted that since the release of DeepSeek-R1 earlier this year, Huawei’s R&D teams had been under pressure to keep pace with customer demands on its Ascend 910B and 910C processors.
“Our customers have raised many issues and expectations they’ve had with Ascend. And they keep giving us great suggestions,” Xu said, adding that the open-source push is a direct response to those challenges.
Joe Green
Breaking Down the Open-Source Commitments
- CANN (Compute Architecture for Neural Networks): Huawei will open interfaces for the compiler and virtual instruction set, while fully open-sourcing other parts of the toolkit. This provides developers with visibility into performance-critical processes without exposing every proprietary element.
- Mind Series Toolchains: All application-level SDKs, debugging tools, and development utilities will be fully open-sourced, giving developers the ability to extend and improve the environment.
- OpenPangu Foundation Models: Huawei confirmed it will make its foundation models publicly available, though details around licensing, training data, and fine-tuning flexibility remain unspecified.
- UB OS Component: To reduce deployment friction, Huawei will open-source its operating system component that manages SuperPod interconnects, allowing integration with existing Linux distributions like openEuler, Ubuntu, or Red Hat Enterprise Linux.
- Framework Support: Compatibility with PyTorch and vLLM is a top priority, making it easier for developers to run familiar workloads and optimize large language model inference on Ascend hardware.
What’s at Stake for Developers
By setting a firm year-end deadline, Huawei is positioning itself alongside other tech giants that have embraced open-source AI, from Meta’s Llama models to Mistral AI’s lightweight frameworks.
For developers, the December release will be the first real test: Is the documentation complete? Are the tools production-ready? Can the community meaningfully contribute? Huawei’s long-term commitment to governance, licensing transparency, and external collaboration will determine whether this is the start of a vibrant ecosystem or just code placed on GitHub without follow-through.
Looking Ahead
The next three months are crucial. Developers and organizations weighing Ascend adoption will need to prepare to evaluate the software the moment it drops. By mid-2026, the market should know whether Huawei’s open-source gamble has paid off in building a global developer community — or whether adoption remains limited to vendor-driven initiatives.
For now, Huawei’s move reflects a growing industry reality: in AI, openness is no longer optional. It’s the entry ticket to developer trust and long-term relevance.
