Chinese AI 'escapes' US chips for the first time.

Zhipu AI has announced a new image generation model trained entirely on Chinese Huawei chips, marking a step forward in its efforts to reduce reliance on US semiconductor technology.

Zhipu AI CEO Zhang Peng speaks at the company's large-scale AI model launch conference in Beijing, China, on March 14, 2024. - Photo: Zhipu AI

According to the SCMP, on January 15, Zhipu AI announced that its GLM-Image imaging model was trained entirely using domestically developed chips from Huawei Technologies, becoming China's first powerful open-source model that does not use any American chips throughout the entire process.

According to Zhipu, the entire GLM-Image training cycle, from data processing and model refinement to the final training run, is performed on Huawei's Ascend Atlas 800T A2 server, using Huawei's own Ascend AI processor.

Zhipu AI, a Beijing-based company that recently completed its IPO in Hong Kong, claims this achievement demonstrates the feasibility of developing advanced multimodal AI models without relying on American semiconductor chips.

Leaders of artificial intelligence company Zhipu AI, semiconductor company Shanghai Iluvatar CoreX, and surgical robotics company Shenzhen Edge Medical attend the listing ceremony at the Hong Kong Stock Exchange (HKEX) on January 8 - Photo: REUTERS

"We hope this result will become a valuable reference for the community, thereby unlocking the potential of domestic computing capabilities," Zhipu AI stated.

Amidst US restrictions on the export of high-end AI chips to China, Beijing is aggressively pursuing a strategy of technological self-reliance, particularly in the field of artificial intelligence – a sector considered a key driver of global economic growth and technological competitiveness.

According to Zhipu, GLM-Image achieved industry-leading scores among open-source models, particularly in its ability to render text in images and generate Chinese characters, a major technical challenge for current image generation models.

Multimodal AI models, capable of simultaneously processing text, speech, images, and video , are being hailed by experts as the "next frontier" of artificial intelligence. This is also the area where global tech giants are fiercely competing.

Zhipu stated that GLM-Image uses a hybrid architecture, combining autoregressive and diffusion models.

This design allows the model to both create high-quality images and process text accurately – an approach previously adopted by Google DeepMind with its advanced multimedia models.

Zhipu AI's move is seen as a signal that China's domestic AI ecosystem is gradually maturing, as the country seeks to develop independently in the face of external technological barriers.

THANH HIEP

Source: https://tuoitre.vn/ai-trung-quoc-lan-dau-thoat-chip-my-20260115213939417.htm