GLM-5.3-Flash

GLM-5.3-Flash is the first natively multimodal model in the GLM-5 family developed by Zhipu AI (Z.ai). It is a Mixture-of-Experts model with 320B total parameters and 18B active parameters per token (288 routed experts, 1 shared, 8 activated per token, 45 layers). It is the first in the series to combine sparse and linear attention in a hybrid architecture, sharply reducing long-context serving cost while preserving long-context quality.

The model introduces Manifold-Constrained Hyper-Connections (mHC) to improve scaling efficiency and an IndexPool mechanism to cut indexer overhead at a 1M-token context length. It was trained on a 30T-token multimodal pre-training corpus. It accepts text, image, video and file input, produces text and code, and supports function calling, structured (JSON) output, context caching, and a thinking mode controlled by the reasoning_effort parameter (low/high/max). Context window up to 1M tokens, maximum output 128K tokens.

Weights are publicly available on Hugging Face (zai-org) under the MIT license; the model can be deployed locally (SGLang, vLLM, TokenSpeed, Transformers, KTransformers, Unsloth) or via the Z.ai API. It targets coding and agentic workflows (including generating PPTX/PDF/DOCX/XLSX files and operating interfaces) at very low inference cost.