GLM-5.3 Flash

GLM-5.3 Flash is Zhipu AI's open-weights multimodal mixture-of-experts model released August 2026, the first open-weight release of the glm5_next architecture. A 320B-parameter checkpoint activates only 18B parameters per token and supports a 1M-token context window. It combines sparse and linear attention with an IndexPool mechanism and Manifold-Constrained Hyper-Connections (mHC), was pre-trained on a 30T-token multimodal corpus, and is licensed under MIT. Zhipu says production serving runs on domestically developed AI accelerators with W8A8 quantization.