Chinese tech company Zhipu AI has achieved a major milestone by training a state-of-the-art image generation model entirely on Huawei’s domestically-made processors. This success demonstrates that Chinese firms can develop competitive artificial intelligence systems without relying on Western hardware. The breakthrough highlights China’s growing capabilities in AI hardware and software development, especially under current restrictions on access to Western chips.

Training a Leading Multimodal Model on Huawei Chips

In a recent announcement, Zhipu AI revealed that it trained its new multimodal model using Huawei’s Ascend Atlas 800T A2 processors. The entire process, from data preprocessing to large-scale training, was carried out using Huawei’s MindSpore AI framework. This marks the first time a cutting-edge multimodal AI model has been fully trained on Chinese-made chips, without any dependence on Western hardware like Nvidia’s GPUs.

The achievement is particularly significant because last year, the U.S. Commerce Department added Zhipu AI to a list of entities deemed to pose security concerns. This restricted the company from purchasing advanced Nvidia GPUs, which are standard for AI training worldwide. As a result, Chinese AI firms have been forced to seek alternatives based on domestic chip architectures. Huawei’s Ascend processors have become a key substitute, enabling Chinese companies to continue their AI development efforts despite export restrictions.

Implications for China’s AI Industry and Strategic Significance

The successful training of this model on Huawei chips shows that Chinese firms can build competitive AI systems domestically. Zhipu AI’s statement emphasized that this proves the feasibility of training high-performance multimodal models on fully domestic hardware and software stacks. This development is seen as a strategic victory, reducing reliance on Western technology and boosting China’s self-sufficiency in AI hardware manufacturing.

In addition to the technical breakthrough, Zhipu AI has made the model accessible through an API, charging a very low fee of 0.1 yuan (about $0.014) per generated image. The model weights are also publicly available on platforms like GitHub, Hugging Face, and ModelScope Community, allowing independent developers and enterprises to deploy the model on their own infrastructure. This move aims to promote widespread adoption and demonstrate the cost-effectiveness of domestically developed AI tools for content creation and marketing.

From a technical standpoint, the model, called GLM-Image, combines a 9-billion-parameter autoregressive component with a 7-billion-parameter diffusion decoder. The autoregressive part helps the model understand instructions and compose images, while the diffusion decoder adds fine details and accurate text rendering. This hybrid architecture addresses challenges in generating complex visual content, such as infographics, presentation slides, and commercial posters.

Performance Benchmarks and Future Prospects

In benchmark tests, GLM-Image achieved impressive results. It scored a Word Accuracy of 0.9116 on the CVTG-2K benchmark, which measures how well models place text across images. It ranked first among open-source models in this test. The model also performed strongly on the LongText-Bench test, where it generated extended text passages. It scored 0.952 for English and 0.979 for Chinese across various scenarios, including signs and labels, showcasing its ability to handle knowledge-intensive visual content.

This achievement not only demonstrates the technical capabilities of Chinese AI developers but also signals a shift towards self-reliance in AI hardware. By successfully training advanced models on domestic chips, Chinese companies are proving that they can compete globally even under export restrictions. The continued development of such models could accelerate AI innovation within China and reduce dependence on Western technology for future projects.

Overall, Zhipu AI’s milestone marks an important step in China’s AI landscape, illustrating that high-quality, large-scale AI models can be built using only domestically available hardware. This progress could shape the future of AI development in the region and inspire other Chinese companies to pursue similar innovations using homegrown technology. As the industry advances, it’s likely we will see more breakthroughs driven by Chinese hardware and software solutions, shaping the global AI market in new ways.

Inspired by

• https://www.computerworld.com/article/4116792/chinese-ai-firm-trains-state-of-the-art-model-entirely-on-huawei-chips-2.html

Sources

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