Zhipu officially launched its large model GLM-5.3-FlashX on September 18. The model's peak inference speed can reach 200 tokens/s, aiming to provide enterprises and developers with a faster and smoother model experience. The API for GLM-5.3-FlashX is now live. According to the company, in order to meet rapidly growing user demand, Zhipu has further increased its investment in infrastructure and inference optimization on top of its existing domestic chip computing power. GLM-5.3-Flash continues to deliver the strongest intelligence level within its size class and offers extremely high cost-performance; this speed boost further enhances its overall competitiveness across intelligence, price and speed.
It is understood that the previously launched GLM-5.3-Flash, before its official release, was tested at scale by developers worldwide under the anonymous model name "Ox Alpha" and received broad recognition. Since launch, its invocation volume has continued to climb, at one point topping the usage rankings on the OpenRouter platform, while simultaneously setting new records for invocation volume on both OpenCode and OpenRouter. Zhipu stated that handling all online invocation requests for GLM-5.3-Flash required a total of 100,000 domestic chips.
Although the specific chip vendor has not been disclosed, industry analysts speculate that it most likely uses Huawei Ascend series chips.
Zhipu's GLM-5.3-FlashX delivers high-efficiency inference
At the cluster level, Zhipu adopted a production-grade EPD disaggregated architecture, splitting multimodal encoding, prompt prefill, and per-token decoding into independently scheduled and independently scaled workload pools. End-to-end service performance has improved threefold compared with the initial baseline, with hardware efficiency and per-token cost now reaching levels comparable to mainstream NVIDIA GPUs. Zhipu stated that this optimization validates domestic chips' ability to efficiently and economically support frontier model inference at scale.