Open-source AI development took centre stage at Huawei Connect 2025 last week, with Huawei laying out implementation timelines and the technical specifics around making its entire AI software stack publicly available by year-end.
The announcements came with context that matters to developers: frank acknowledgement of past friction, specific commitments about what components will be released, and details about how the software will integrate with existing workflows and operating systems.
Developer friction acknowledged
Eric Xu, Huawei’s Deputy Chairman and Rotating Chairman, opened his keynote with unusual candour about challenges developers have faced with Ascend infrastructure. Referencing the impact of DeepSeek-R1’s release earlier this year, Xu noted: “Between January and April 30, our AI R&D teams worked closely to make sure that the inference capabilities of our Ascend 910B and 910C chips can keep up with customer needs.”
Following customer feedback sessions, Xu stated: “Our customers have raised many issues and expectations they’ve had with Ascend. And they keep giving us great suggestions.”
Acknowledgement of developer pain points provided context for the comprehensive open-source commitments announced at the August 5, 2025 Ascend Computing Industry Development Summit and reinforced by Xu at Huawei Connect.
For developers who have struggled with Ascend tooling, documentation, or ecosystem maturity, the frank assessment signals awareness of gaps between the platform’s technical capabilities and its practical usability. The open-source strategy appears designed directly to address these friction points by enabling community contributions, transparency, and external improvements.
CANN: Compiler and virtual instruction set details
The most technically significant commitment involves CANN (Compute Architecture for Neural Networks), Huawei’s foundational toolkit that sits between AI frameworks and Ascend hardware. At the August summit, Xu specified: “For CANN, we will open interfaces for the compiler and virtual instruction set, and fully open-source other software.”
The tiered approach distinguishes between components receiving full open-source treatment versus those where Huawei will provide open interfaces with potentially proprietary implementations. The compiler and virtual instruction set – important translation layers that convert high-level code into hardware-executable instructions – will have open interfaces. This means developers can understand and potentially optimise how their code gets compiled for Ascend processors, even if the compiler implementation itself remains partially closed.
The distinction matters for performance tuning. Developers need visibility into compilation processes when working on latency-sensitive applications or are trying to extract maximum efficiency from hardware. Open interfaces provide that visibility; full open-source would additionally enable replacing or modifying the compiler itself. Huawei’s approach offers transparency for optimisation yet retains some proprietary elements.
The timeline remains firm: “We will go open source and open access with CANN (based on existing Ascend 910B/910C design) by December 31, 2025.” The specification of “based on existing Ascend 910B/910C design” clarifies that the open-source release will reflect current-generation hardware rather than future chip architectures.
Mind series: Application enablement kits and toolchains
Beyond the foundational CANN layer, Huawei committed to open-sourcing what developers interact with daily: “For our Mind series application enablement kits and toolchains, we will go fully open-source by December 31, 2025,” Xu said at Huawei Connect, reinforcing the commitment made at the Ascend Computing Industry Development Summit on August 5, 2025.
The Mind series encompasses the practical development environment – the SDKs, libraries, debugging tools, profilers, and utilities that developers use when building AI applications. Unlike CANN’s tiered approach with open interfaces for some components, the Mind series sees a blanket commitment to full open-source.
This means the entire application layer toolchain becomes inspect-able, modifiable, and community-extensible. Debugging tools could be enhanced with needed functionality, libraries can be optimised for specific use cases, and utilities can be wrapped in more ergonomic interfaces. In short, the development ecosystem will evolve through community contributions rather than standing on vendor updates.
However, the announcement didn’t specify which tools specifically comprise the Mind series, which programming languages they support, or how comprehensive the documentation is to be. Developers evaluating whether to invest time in the platform will need to assess toolchain completeness once the December release arrives.
OpenPangu foundation models
Huawei has also committed to “fully open-source our openPangu foundation models.” This positions Huawei in the open-source foundation model space alongside Meta’s Llama series, Mistral AI’s offerings, and various other initiatives that lean into community involvement.
The announcement provided no specifics about openPangu capabilities, parameter counts, training data, or licensing terms. Foundation model open-sourcing raises questions beyond licensing, and what restrictions will exist on commercial use. What datasets were used for training, and what biases or limitations does each model exhibit? Can the model be fine-tuned and redistributed? These issues have yet to be resolved, at least publicly.
For developers, open-source foundation models provide starting points for domain-specific applications without requiring the massive computational resources needed for training from scratch. However, model quality, licensing flexibility, and available documentation determine practical utility. The December release will reveal whether openPangu models represent competitive alternatives to established open-source options.
Operating system integration flexibility
One practical implementation detail that emerged at Huawei Connect 2025 addresses a common barrier to adopting new AI infrastructure: operating system compatibility. Huawei announced that “Huawei has made the entire UB OS Component open-source, so that its code can be integrated into upstream open-source OS communities like openEuler.”
The integration approach offers unusual flexibility. According to the announcements: “Users can integrate part or all of the UB OS Component’s source code into their existing OSes, to support independent iteration and version maintenance. Users can also embed the entire component into their existing OSes as a plug-in to ensure it can evolve in-step with open-source communities.”
Modular design means organisations running Ubuntu, Red Hat Enterprise Linux, or other distros aren’t forced to migrate to a Huawei-specific operating systems. The UB OS Component – which handles SuperPod interconnect management at the operating system level – can be integrated into existing environments. For developers and system administrators, this lowers deployment friction significantly.
However, flexibility comes with responsibility. Organisations choosing to integrate UB OS Component source code into their own systems become responsible for testing, maintenance, and updates. Huawei is providing the component as open-source rather than as a supported product for arbitrary Linux distributions. The approach works well for organisations with strong Linux expertise; it may prove challenging for those expecting turnkey vendor support.
Framework compatibility strategy
Perhaps the most important factor for developer adoption is compatibility with existing AI frameworks. Rather than forcing developers to abandon familiar tools, Huawei is building integration layers. According to Huawei, it “has been prioritising support for open-source communities like PyTorch and vLLM to help developers independently innovate.” PyTorch compatibility is particularly significant given that framework’s dominance in AI research and production deployments. If developers can write standard PyTorch code that executes efficiently on Ascend hardware without extensive modifications, the barrier to experimentation drops substantially.
Organisations could evaluate Ascend infrastructure using minimally-tweaked existing codebases rather than requiring rewrites. The vLLM integration targets a specific high-demand use case: optimised large language model inference. As organisations deploy LLM-based applications, inference performance and cost become important factors.
Native vLLM support suggests Huawei is addressing practical deployment concerns rather than just research capabilities. However, the announcements didn’t detail the completeness of any integration. Partial PyTorch compatibility that requires workarounds for certain operations or delivers suboptimal performance may prove more frustrating than existing alternatives. The quality of framework integrations will determine whether they genuinely lower adoption barriers or simply create new categories of compatibility issues.
December 31 deadline and what follows
The December 31, 2025 timeline for open-sourcing CANN, Mind series, and openPangu models is approximately three months away. The near-term deadline suggests substantial preparation work is already complete: code has been cleaned of internal dependencies, documentation is being written, licensing terms are being finalised, and repository infrastructure is being established.
Initial release quality will largely determine community response. Open-source projects that arrive with incomplete documentation, limited examples, missing features, or immature tooling often fail to attract contributors regardless of underlying technical merit.
Developers evaluating unfamiliar platforms need comprehensive learning resources, working examples, and clear paths from “Hello World” to production deployment. The December release represents a beginning rather than a culmination.
Successful open-source projects require sustained investment beyond initial code publication. Community management, issue triage, pull request review and merge, documentation maintenance, and roadmap coordination all demand ongoing resources. Whether Huawei commits to multi-year community support will determine whether the platform develops an active contributor base or becomes abandoned code with public repositories but minimal development activity.
What remains unspecified
Despite the specific commitments and timelines, several important details about open-source AI development on Ascend remain undefined. Licence selection will fundamentally affect how developers and organisations can use the software. Permissive licences like Apache 2.0 or MIT enable commercial use with minimal restrictions and allow proprietary derivatives.
Copyleft licenses like GPL require derivative works to also be open-sourced, which affects traditional models of commercial product development. Huawei hasn’t specified under which licences the December releases will be. Overall governance structures for the open-source projects are equally unclear.
Will there be an independent foundation overseeing development? Will Huawei accept external maintainers with commit privileges? How will feature priorities and roadmap decisions be made? Will there be a transparent process for accepting community contributions?
Governance questions often determine whether projects attract genuine external participation or remain vendor-controlled initiatives with public code but limited community influence.
Developer evaluation timeline
For developers and organisations considering investment in Huawei’s open-source AI development platform, the next three months provide time for preparation and evaluation. Organisations can assess their requirements, evaluate whether Ascend hardware specifications match their workload characteristics, and prepare teams for potential platform adoption.
The December 31 release will provide concrete materials for hands-on evaluation: actual code to review, documentation to assess, examples to test, and toolchains to experiment with. The weeks following release will reveal community response – whether external developers file issues, contribute improvements, and begin building the ecosystem resources that make platforms increasingly capable.
By mid-2026, patterns should have emerged about whether Huawei’s open-source AI development strategy is succeeding in building an active community around Ascend infrastructure or whether the platform remains primarily a vendor-led initiative with limited external participation.
For developers, a six-month window from December 2025 through to around mid-2026 will be an evaluation period for determining whether this open-source platform warrants serious investment of time and resources.
See also: Inside Huawei’s plan to make thousands of AI chips think like one computer
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and co-located with other leading technology events. Click here for more information.
AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
