Zhipu AI Open-Sources GLM-5.3-Flash Multimodal Model with 320B Total Parameters and 18B Active Parameters

Zhipu AI today launched and open-sourced the GLM-5.3-Flash (320B-A18B) model, the first natively multimodal model in the GLM-5 series, featuring 320 billion total parameters and 18 billion activated parameters. Prior to its official release, the model was tested under the anonymous name Ox-Alpha and became the most frequently invoked model on both OpenCode and OpenRouter platforms, with all inference traffic powered by domestically developed Chinese chips. The model scored 57 on the Artificial Analysis Intelligence Index (AA Composite Intelligence Index), matching Anthropic’s Claude Opus 4.8. On Zhipu AI’s proprietary Z.ai CodeBench benchmark, its coding capabilities are on par with Claude Opus 4.8. Priced at one-tenth the cost of GLM-5.3—and temporarily discounted to one-twentieth during the promotional period—GLM-5.3-Flash costs just one-fortieth of Claude Opus 4.8. Designed specifically for low-cost deployment, it has a total parameter count comparable to GLM-4.5 but nearly halves both activated parameters and model depth. Trained on 30 trillion tokens of multimodal pretraining data, the model employs a hybrid architecture combining sparse attention and linear attention, enhanced with manifold-constrained hyperconnectivity (mHC) technology. GLM-5.3-Flash natively integrates visual capabilities into its code generation pipeline, enabling coordinated task execution across code environments, web browsers, and graphical user interfaces. Target applications include office productivity, financial research, and professional document processing.