Technology

Chinese AI startup launches image model trained entirely on Huawei chips

Wednesday 14 - 18:50

By: Dakir Madiha

Chinese AI startup launches image model trained entirely on Huawei chips

A Chinese artificial intelligence startup has unveiled an advanced open-source image generation model, marking a pivotal advancement in China's drive for technological self-reliance amid ongoing U.S. export restrictions. Zhipu AI introduced GLM-Image, the first top-tier multimodal AI model fully trained on domestically produced hardware, specifically Huawei's Ascend Atlas 800T A2 chips and its MindSpore machine learning framework. This breakthrough underscores Beijing's efforts to develop alternatives to Western components, particularly after Zhipu was added to the U.S. Commerce Department's entity list in January 2025 for alleged contributions to China's military modernization through AI development.

The announcement propelled Zhipu shares, traded under Knowledge Atlas Technology JSC Ltd., up more than 17 percent, fueling a broader rally in Chinese chipmaking stocks. In a statement, Zhipu highlighted the model's feasibility for high-performance generative multimodal training on a fully local computing stack, offering it as a valuable reference for the community to explore national computing potential. GLM-Image features a hybrid architecture with a 9-billion-parameter autoregressive model paired with a 7-billion-parameter diffusion decoder, excelling in benchmarks for text rendering accuracy. It topped the CVTG-2K benchmark with a word accuracy score of 0.9116 and led LongText-Bench with scores of 0.952 in English and 0.979 in Chinese among open-source alternatives.

This launch comes just a week after Zhipu's Hong Kong Stock Exchange debut, raising 4.17 billion Hong Kong dollars, 70 percent earmarked for research and development making it the first of China's leading "AI tigers" to go public. The timing coincides with U.S. approval under the Trump administration for limited exports of NVIDIA's H200 chips to China, capped at 50 percent of domestic U.S. sales, though Chinese customs authorities have restricted purchases to narrow cases like university research collaborations. Zhipu has made GLM-Image accessible via an API at 0.1 yuan (about $0.014) per generated image, appealing to businesses creating text-rich marketing materials and visuals, while releasing model weights on GitHub, Hugging Face, and ModelScope for independent deployment.