Zhipu AI Advances China's Chip Independence With Huawei-Trained Model
- tech360.tv
- 3 hours ago
- 2 min read
Chinese artificial intelligence firm Zhipu AI announced its new image generation model was trained on chips from Huawei Technologies, making it the first powerful open-source model developed on an entirely domestic training stack, marking a significant step towards domestic technology independence. This development demonstrates the feasibility of creating powerful multimodal models without relying on United States semiconductors.
The Beijing-based company achieved this milestone as China pushes for self-reliance in its AI industry. This push comes amid US export restrictions on cutting-edge chips.
Zhipu stated that GLM-Image's entire training process, from data preparation to the final run, used Huawei's Ascend Atlas 800T A2 server. This server incorporates Huawei's in-house Ascend AI processors, and MindSpore, Huawei's machine learning framework.
A Zhipu spokesperson said, "We hope this can provide valuable reference for the community to explore the potential of domestic computing power." Industry experts consider powerful multimodal AI models, capable of processing text, voice, image, and video, to be the next frontier in AI.
Zhipu's model features a hybrid architecture, combining autoregressive and diffusion elements. This design enables multimodal capabilities, a pioneering feature seen in models such as Google DeepMind's Nano Banana Pro, which accurately generates images and text.
Zhipu claimed the model achieved industry-leading scores among open-source models for text rendering. It was particularly effective at Chinese character generation.
However, self-reported scores indicated the GLM-Image model still lagged behind ByteDance's proprietary image generation model, Seedream 4.5. ByteDance has not disclosed the chips used to train its model.
Huawei's Ascend chips have proven effective for training smaller models like Zhipu's GLM-Image. Their efficacy for training the company's flagship large language models, such as the next-generation GLM-5, remains to be determined, according to a person familiar with the matter.
Nvidia chips currently dominate cutting-edge model development in China. Few Chinese AI models publicly disclosed as trained on domestic chips are limited to second-tier model developers like iFlytek.
The US government blacklisted iFlytek in 2019, and Zhipu was added to the same export control list in Jan. last year. Beijing has encouraged Chinese technology firms to adopt domestic machine learning frameworks, including Huawei's MindSpore, over US alternatives.
Despite Beijing's encouragement, MindSpore's adoption remains limited compared to more established US frameworks. These frameworks include TensorFlow and PyTorch, developed by Google and Meta Platforms, respectively, according to data from various open-source developer platforms.
Zhipu AI trained its new GLM-Image model entirely on Huawei Technologies chips and software.
This marks the first powerful open-source model developed on an entirely domestic Chinese training stack.
The achievement demonstrates China's progress towards self-reliance in AI development, reducing dependence on US semiconductors.
Source: SCMP
