Zhipu launches the native multimodal large model GLM-5.3-Flash, running 'near Opus' capabilities on domestic chips at just one-tenth the cost. According to http://Z.ai's official website, Zhipu has introduced the native multimodal large model GLM-5.3-Flash, featuring a total of 320 billion parameters and 18 billion active parameters. It significantly outperforms GLM-5.2 in encoding and agent benchmarks and approaches Claude Opus 4.8, with costs reduced to about one-tenth. The model leverages a hybrid architecture of sparse and linear attention, Manifold-Constrained Hyper-Connections, and 30 trillion multimodal datasets, reducing long-context reasoning costs to about one-third of GLM-5.3. On the Artificial Analysis Intelligence Index v4.1.1, it achieved a performance score of 57 at approximately $0.045/task. The official statement claims that during anonymous testing on OpenCode and OpenRouter under the ox-alpha tag, it became the hottest model within a week and has achieved inference efficiency close to NVIDIA GPUs on large-scale domestic AI chip clusters.