Chinese AI firm Zhipu AI has announced a significant milestone in the advancement of artificial intelligence: the training of a new large language model,GLM-Image,entirely on domestically produced Huawei hardware. This achievement marks the frist instance of an advanced AI model being built using a complete Chinese hardware stack, perhaps lessening reliance on Western technology and reshaping the global AI landscape.
Zhipu AI and the GLM-Image Model
Zhipu AI, also known as Z.ai, operates a chatbot accessible at chat.z.ai and offers a suite of models under the General Language Model (GLM) umbrella. On Wednesday, the company unveiled GLM-image, a multimodal model capable of generating both images and text. According to Zhipu, GLM-Image employs a novel “autoregressive + diffusion decoder” hybrid architecture, representing an advancement beyond models like Nano Banana Pro in image generation capabilities.
Huawei Hardware at the Core
A key aspect of this development is Zhipu AI’s reliance on Huawei’s hardware infrastructure. the model was trained using the Ascend Atlas 800T A2 server, a powerful system designed for AI workloads.Specifically, the Atlas 800T A2 houses four Kunpeng 920 processors, each containing either 64 or 48 cores, and utilizes Huawei’s Ascend 910 AI processors to accelerate computations.Huawei’s Kunpeng processors are based on the Arm architecture,which they’ve designed themselves.
Ascend Processors: A Competitive Edge
Huawei has been aggressively developing its Ascend line of processors to compete with industry leaders like NVIDIA. The latest iteration, the Ascend 910C (introduced in 2025), boasts approximately 800 TFLOPS of computing power at FP16 precision – a figure that Huawei claims represents 80% of the performance offered by NVIDIA’s 2022 H100 chip. While autonomous verification is necessary, these figures highlight Huawei’s increasing capabilities in the AI hardware space.
GLM-Image: Architectural Details
Zhipu AI provided insights into the architecture of GLM-Image via Hugging Face. The model comprises two critical components:
- Autoregressive Generator: A 9 billion parameter model initially based on GLM-4-9B-0414. It’s adapted to handle visual tokens, first creating a compact 256-token encoding that expands to 1K-4K tokens, ultimately generating high-resolution images (1K-2K pixels).
- Diffusion decoder: A 7 billion parameter decoder employing a DiT architecture for image decoding within the latent space. The inclusion of a Glyph Encoder text module significantly improves the model’s ability to accurately render text within generated images.
Implications for China and the Global AI Race
Zhipu AI emphasizes that the entire development process, “from data preprocessing to large-scale training,” was conducted using the Huawei Atlas server. This demonstrates, according to the company, the “feasibility of training cutting-edge models on a domestically produced full-stack computing platform.” However, Zhipu has not yet disclosed the precise number of servers or accelerators used, or the training duration, which makes a direct performance comparison difficult.
Despite these unknowns, the news carries significant weight. Even if the training process wasn’t dramatically faster than alternatives, the ability to develop a sophisticated AI model using only indigenous technology is a strategic advantage for China. It reduces dependence on foreign suppliers and could insulate the nation from geopolitical pressures and export controls. This development is notably relevant given the recently announced US export controls targeting advanced GPUs, which will subject all sales of certain GPUs to Chinese buyers to thorough vetting.
A Future of Niche Models?
Amidst predictions that future AI models will increasingly specialize in narrow domains, China’s ability to independently develop models like GLM-Image is a crucial step. If China can reliably produce these models without reliance on NVIDIA or AMD hardware, it poses a direct challenge to the revenue streams of those leading chip designers.
Broader Geopolitical Considerations
The development also taps into broader geopolitical concerns highlighted by the australian Strategic Policy Institute (ASPI). ASPI has warned about China’s potential to use AI to spread its cultural values and political influence globally. The development of independent AI capabilities could further strengthen China’s soft power and its ability to shape the global technological landscape.
GLM-Image is released as an open-source model, making it freely available for research and development. While questions remain about the efficiency and cost-effectiveness of building AI models on Huawei hardware, this achievement signals a shift toward greater self-reliance in China’s AI ambitions and introduces a new dynamic into the global AI competition.
