China’s Z.AI has released a major open-source image model trained entirely on Huawei chips.
It uses a hybrid autoregressive diffusion design that improves text accuracy and spatial control.
The release signals China’s push for AI self-reliance without American GPUs.
Chinese artificial intelligence company Z.AI on Wednesday released an open-source image generation model fully trained on Huawei processors. This is the first time a major AI model has completed its full training cycle without relying on US hardware.
The move highlights a potential long-term challenge to Nvidia’s dominance in AI chips, as it shows that one of China’s top AI companies can train large models without relying on US-made GPUs.
The model is already available for download on Hugging Face and delivers good – but not impressive by today’s standards – results in terms of aesthetics and coherent text, and shows excellent spatial awareness based on our initial quick tests.
Image generated with Z.AI’s new model.
The Beijing-based company, which raised $558 million in its Hong Kong IPO last week, trained the model, called GLM-Image, on Huawei’s Ascend Atlas 800T A2 servers using the MindSpore framework.
“We hope this can be a valuable reference for the community to explore the potential of domestic computing power,” Z.AI said in a statement shared with the South China Morning Mail.
Introducing GLM-Image: a new milestone in open-source image generation.
GLM-Image uses a hybrid autoregressive plus diffusion architecture, combining strong global semantic understanding with high-fidelity visual detail. It matches regular diffusion models in overall quality… pic.twitter.com/cjtUYRkge5
GLM-Image combines autoregressive and diffusion techniques in a hybrid architecture with a total of 16 billion parameters. The autoregressive component, based on Z.AI’s GLM-4 language model, handles instruction understanding and image composition, while a diffusion decoder refines fine details. This approach mirrors the techniques used by OpenAI’s latest image generation model gpt-image-1.5, which has demonstrated superior text rendering and fast compliance compared to pure diffusion models such as Stable Diffusion.
Diffusion models create images by starting with random visual noise and slowly refining them into an image, while autoregressive models build images step by step, predicting each component based on what came before it. Diffusion is good at overall realism but can struggle with precise details such as text or layout, while autoregressive models excel at structure and instruction following. Currently, diffusion is the king technique among open-source AI image generators.
New hybrid systems combine both approaches, using autoregressive generation to plan the image and diffusion to improve the final result.
Image: Z.AI
The release carries weight for Z.AI, which Washington blacklisted in 2025 over alleged ties to the Chinese military. That designation excluded the company from Nvidia’s H100 and A100 processors. Now Z.AI has proven that blacklisted companies can still produce competitive AI systems using domestic hardware, a development that Beijing has long been trying to demonstrate.
Just after Z.AI’s announcement, Reuters reported that Chinese customs authorities had ordered agents to prevent Nvidia H200 chips from entering the country. Government officials called tech companies to meetings where they were told not to buy the chips unless necessary. Sources said the wording was strict enough to constitute “effectively a ban for the time being.”
Beijing appears to be signaling that Chinese AI labs can build capable models without American silicon, reducing the urgency for Chinese companies to stock up on Nvidia hardware. The H200, which delivers roughly six times the performance of the H20 chip that Beijing blocked last August, had generated orders from Chinese companies for more than two million units at $27,000 each.
Analysts at Georgetown’s Center for Security and Emerging Technologies have noted that China’s chip strategy relies on offsetting lower per-chip performance with huge clusters of Huawei processors. The approach works, but requires more hardware, more power and more engineering effort.
“One of the key limitations of this strategy is China’s ability to produce enough chips domestically to fill and keep up with the capacity gap,” senior research analyst Hanna Dohmen told me. CNBC in November.
According to Huawei’s own roadmap, its next-generation chip will actually be worse than its current flagship in terms of raw power by 2026. But such assessments may underestimate what Chinese labs can achieve through algorithmic efficiency, as DeepSeek demonstrated by training competing models with fewer chips through GPU optimization at the assembly level.
Source: Council on Foreign Relations
Z.AI’s GLM-Image achieved leading benchmark scores among open-source models for text rendering and Chinese character generation, according to the company’s technical report. Those without the right hardware can also try it online with API access for $0.014 per generated image, or via a free Hugging Face Space maintained by Z.AI.
Z.AI became the first of China’s “AI Tigers,” a group of startups building large language models to rival OpenAI and Anthropic, to go public. The stock has risen about 80% since listing, reflecting investor enthusiasm for Chinese AI companies like DeepSeek or Alibaba amid China’s domestic chip ambitions.
Huawei, meanwhile, is preparing to sharply increase production of its Ascend processors this year. The company’s presence at AI conferences across China has become more prominent as it tries to position itself as the backbone of a national AI infrastructure that no longer relies on Santa Clara.
