Zhipu released GLM-5.3 Flash, a lightweight flagship model positioned to rival DeepSeek V4 Flash. The company says inference is served by a cluster using more than 100,000 domestically produced accelerator cards. Compute suppliers are reported to include Huawei, Moore Threads and Hygon; Zhipu declined to comment. In technical documentation Zhipu said the cluster's "hardware efficiency and per-token cost are already comparable to mainstream NVIDIA GPUs," but it did not specify card models.