AI 导读
智谱于 8 月 26 日开源 GLM-5.3-Flash，这是 GLM-5 系列首个原生多模态模型，其线上流量全部运行在 10 万张国产芯片上。发布前，匿名模型 Ox Alpha 在 OpenRouter 和 OpenCode 上处理了 62T tokens，同样使用中国硬件，据报道供应商可能包括华为、海光与摩尔线程。该模型在 Artificial Analysis 上得分 57，与 Claude Opus 4.8 持平，成本为 GLM-5.3 的十分之一。
正文
Zhipu open-sourced GLM-5.3-Flash on Aug. 26, the GLM-5 series’ first native multimodal model. Its online traffic runs entirely on 100,000 domestic chips.
Before launch, the anonymous “Ox Alpha” model processed 62T tokens on OpenRouter and OpenCode—also using Chinese hardware. Reported suppliers may include Huawei, Hygon and Moore Threads.
It scored 57 on Artificial Analysis, tying Claude Opus 4.8, while costing one-tenth of GLM-5.3. Open model, domestic compute, global demand.
来源：@thexpin · x.com