Z.ai Debuts China's First Domestically Trained Multimodal AI Model Using Huawei Chips

Chinese AI firm Knowledge Atlas Technology (Z.ai) has launched GLM-Image, positioning it as China's first multimodal AI model entirely trained using domestically produced semiconductor technology. The open-source model, hosted on Hugging Face and GitHub, leverages Huawei's Ascend processors—a strategic shift accelerated by U.S. restrictions on advanced AI chip exports to China.

GLM-Image combines image recognition with natural language processing, enabling applications from visual question answering to contextual image generation. Benchmark details haven't been disclosed, but the model aligns with China's "Made in China 2025" semiconductor roadmap, which aims to eliminate foreign dependency in critical technologies. Huawei's Ascend chips now power the entire training pipeline, from data preprocessing to inference optimization.

The release comes amid tightening U.S. export controls. Last week, the Commerce Department approved limited Nvidia H200 chip sales to China under case-by-case reviews—but capped volumes at 50% of U.S. customer allocations. Z.ai's achievement demonstrates China's accelerating progress in building sovereign AI infrastructure. Huawei's Ascend series, while less efficient than Nvidia's latest offerings, now handles full-stack model training without foreign components.

For China's AI ecosystem, GLM-Image offers three strategic advantages:

1. Supply chain insulation: Eliminates risks from geopolitical disruptions
2. Cost control: Domestic chips avoid import tariffs and logistics premiums
3. Regulatory compliance: Meets China's data sovereignty requirements

Open-sourcing the model could accelerate adoption across Chinese academia and enterprise sectors. However, performance gaps remain: Huawei's chips reportedly consume 40% more power than Nvidia equivalents for similar tasks, and software tooling lags behind CUDA's maturity.

This milestone pressures Western AI leaders as China's domestic chip production scales. Semiconductor Manufacturing International Corporation (SMIC) recently achieved 5nm Ascend wafer yields, while Huawei plans to double AI chip output in 2026. With GLM-Image, Z.ai proves that viable alternatives to U.S.-dependent AI pipelines are now operational—potentially reshaping global AI supply chains within two years.