Chinese tech company Zhipu AI has successfully trained a cutting-edge image generation model entirely on Huawei processors. This marks a significant milestone for Chinese firms in developing competitive AI systems without relying on Western hardware, especially Nvidia’s GPUs. The achievement highlights China’s growing ability to build advanced AI tools with domestically made chips amidst ongoing global supply restrictions.

Training the Model on Huawei Hardware

Zhipu AI used Huawei’s Ascend Atlas 800T A2 processors to train its new multimodal AI model, called GLM-Image. The company utilized Huawei’s MindSpore AI framework to manage the entire training process—from preparing data to running large-scale computations. This is the first time a state-of-the-art multimodal model has been fully trained on Chinese-made chips, according to Zhipu.

This development is especially important because last year, the US Commerce Department added Zhipu AI to a list of entities deemed to pose national security concerns. As a result, the company was cut off from Nvidia’s H100 and A100 GPUs, which are standard hardware for training advanced AI models. This restriction pushed Chinese companies to find alternatives using domestic chip architectures like Huawei’s Ascend processors.

Implications for China’s AI Industry

Huawei’s chips have become the main hardware option for Chinese AI firms restricted from buying Western GPUs. The successful training of GLM-Image on Huawei chips demonstrates that China can develop high-performance AI models without access to Nvidia’s hardware. This could accelerate the country’s efforts to become self-sufficient in AI technology and reduce reliance on foreign suppliers.

Zhipu AI’s collaboration with Huawei on GLM-Image shows how domestic companies are adapting. The company has made the model available through an API at a low cost—about 0.1 yuan (roughly $0.014) per image generated. It also released the model weights on platforms like GitHub and Hugging Face, allowing other developers to deploy the model independently.

This move makes GLM-Image an affordable choice for businesses needing large amounts of visual content, such as marketing teams or presentation creators. The focus on cost-effectiveness and accessibility could help Chinese firms compete both locally and internationally in AI-powered visual generation.

Technical Highlights and Performance

GLM-Image combines a 9-billion-parameter autoregressive part with a 7-billion-parameter diffusion decoder. The autoregressive component helps understand instructions and create overall images, while the diffusion decoder adds detailed textures and accurate text rendering. This hybrid architecture is designed to handle complex visual tasks that require both semantic understanding and precise text placement.

According to Zhipu, the model performed very well on several benchmarks. It achieved a Word Accuracy score of 0.9116 on CVTG-2K, the highest among open-source models, which measures how accurately the model places and recognizes text in images. It also scored highly on LongText-Bench, with 0.952 for English and 0.979 for Chinese, demonstrating its ability to generate extended text in different languages across various scenarios.

This technical success proves that Chinese companies can develop powerful AI models entirely on domestically produced hardware. It also underscores Huawei’s role as a key hardware provider for China’s AI ambitions, especially as Western restrictions tighten.

Overall, Zhipu AI’s achievement shows that China is making significant strides in AI development. By training advanced models on Huawei chips, Chinese firms can continue innovating despite external challenges. This could reshape the landscape of AI research and deployment in the region, fostering greater independence and technological resilience.

Inspired by

• https://www.infoworld.com/article/4116787/chinese-ai-firm-trains-state-of-the-art-model-entirely-on-huawei-chips.html

Sources

• github.com
• federalregister.gov
• bigmodel.cn
• huggingface.co
• modelscope.cn