Chinese outfit Zhipu AI claims it trained a new model entirely using Huawei hardware, and that it\u2019s the first company to build an advanced model entirely on Chinese hardware.<\/p>\n
Zhipu, which styles itself Z.ai and runs a chatbot<\/a> at that address, offers several models named General Language Model (GLM). On Wednesday the company announced<\/a> GLM-Image, that it says employs \u201can independently developed \u2018autoregressive + diffusion decoder\u2019 hybrid architecture, which enables the joint generation of image and language models.\u201d represents an important advance on the Nano Banana Pro image-generating AI.<\/p>\n The post also states that Z.ai developed the model using the Ascend Atlas 800T A2, a Huawei server<\/a> that can run four Kunpeng 920 processors packing either 64 or 48 cores. Huawei\u2019s processors use Arm cores of its own design.<\/p>\n The servers also use Huawei\u2019s Ascend 910 AI processors.<\/p>\n The most recent Ascend model is 2025\u2019s 910C, which Huawei claims<\/a> \u201ccan achieve around 800 TFLOPS of computing power per card at FP16 precision, which is approximately 80% of the computing power of NVIDIA\u2019s H100 chip (launched in 2022).\u201d<\/p>\n On model-mart Hugging Face, Zhipu describes<\/a> GLM-Image\u2019s architecture as comprising two elements:<\/p>\n Autoregressive generator: a 9B-parameter model initialized from GLM-4-9B-0414, with an expanded vocabulary to incorporate visual tokens. The model first generates a compact encoding of approximately 256 tokens, then expands to 1K\u20134K tokens, corresponding to 1K\u20132K high-resolution image outputs.<\/p>\n Diffusion Decoder: a 7B-parameter decoder based on a single-stream DiT architecture for latent-space image decoding. It is equipped with a Glyph Encoder text module, significantly improving accurate text rendering within images.<\/p>\n The company says \u201cthe entire process from data preprocessing to large-scale training\u201d took place using that Atlas server, and that the model\u2019s debut therefore proves \u201cthe feasibility of training cutting-edge models on a domestically produced full-stack computing platform.\u201d<\/p>\n And in some ways it does. But Zhipu hasn\u2019t revealed how many servers or accelerators it used to create GLM-image, and how quickly they did the job.<\/p>\n The company can therefore point to having developed a model with local tech \u2013 sophistry that ignores Arm\u2019s contribution to Kunpeng \u2013 but hasn\u2019t offered any hints about whether Huawei\u2019s hardware did it at a speed or price that means the rest of the world needs to take notice because China has stolen a march.<\/p>\n Even if Zhipu\u2019s rig chugged along at modest speeds, news of an all-Chinese model remains notable given pundits\u2019 predictions that many future models will be smallish affairs dedicated to niche domains. If China now has the capacity to make such models without hardware from Nvidia or AMD, that\u2019s a threat to those chip design firms\u2019 future revenue.<\/p>\n Another threat to the two GPU giants is the strict export controls<\/a>, announced yesterday, that mean Washington will assess every application to sell certain GPUs to Chinese buyers.<\/p>\n