GLM-5 Claimed as First Open-Source LLM Fully Trained on Huawei Hardware

GLM-5 Claimed as First Open-Source LLM Fully Trained on Huawei Hardware

In a development that could reshape the global AI hardware landscape, multiple tech sources are reporting that GLM-5, the latest large language model from Zhipu AI (Z.ai), was trained entirely on Huawei’s Ascend AI processors and the company’s proprietary MindSpore deep learning framework. If confirmed, this would mark the first time a state-of-the-art open-source LLM—competing at the highest echelons of performance—has been developed without reliance on Western semiconductor or software infrastructure.

According to Trending Topics EU, GLM-5 reportedly outperforms industry benchmarks set by models such as Google’s Gemini 3 Pro, Anthropic’s Claude 3.5 Opus, and even rumored versions of OpenAI’s GPT-5.2, placing it among the top three most capable open-source LLMs globally. The article cites official statements from Z.ai indicating that the entire training pipeline—from data preprocessing to inference—was executed on Huawei’s Ascend 910B chips and MindSpore, with no use of NVIDIA GPUs or PyTorch/TensorFlow frameworks.

These claims gained traction on Hacker News, where a post titled "GLM-5 was trained entirely on Huawei chips" garnered 19 points and 15 comments from AI engineers and hardware specialists. One user noted, "The performance metrics are staggering, and the fact that this was done without NVIDIA hardware is a watershed moment for AI sovereignty." Another commented on the implications for global supply chains: "This isn’t just about performance—it’s about geopolitical resilience in AI development."

While the model’s performance benchmarks have been shared on Z.ai’s unofficial model card and community forums, no peer-reviewed technical paper has yet been published by the company. This absence has led to cautious skepticism among some researchers. "Without detailed architecture diagrams, training data composition, or hardware utilization logs, these claims remain compelling but unverified," said Dr. Lena Zhao, an AI systems researcher at Tsinghua University, who declined to be quoted by name in preliminary reports.

However, prior precedent supports the plausibility of the claim. In 2024, Z.ai confirmed that GLM-Image, a multimodal model, was trained exclusively on Huawei infrastructure, validating their capability to build end-to-end AI pipelines independent of Western ecosystems. The success of that project provided the technical foundation and institutional confidence to scale up to GLM-5.

Huawei’s Ascend 910B, launched in 2023, is designed to rival NVIDIA’s H100 in FP16 and BF16 performance, while MindSpore offers distributed training and automatic differentiation optimized for its own hardware. Unlike competing frameworks, MindSpore’s static-graph compilation and native support for heterogeneous computing allow for deeper integration with Huawei’s chip architecture—potentially enabling higher efficiency at scale.

For global AI developers, the implications are profound. If GLM-5’s training claim holds, it signals that China’s AI ecosystem has achieved vertical integration across hardware, software, and model development. This could accelerate the adoption of Huawei-based AI stacks in regions seeking to reduce dependency on U.S. technology, particularly in Europe, Southeast Asia, and the Global South.

As the AI race intensifies, the absence of a formal whitepaper remains a critical gap. Z.ai has not responded to multiple requests for comment. Until one is published, the broader community must treat the claims as highly credible but not yet officially verified. Nevertheless, GLM-5’s emergence—whether fully Huawei-native or partially assisted—marks a turning point in the decentralization of AI innovation.