GLM-5.3-Flash combines multimodal input with a smaller active model

Z.ai describes GLM-5.3-Flash as a natively multimodal model designed for coding, tool use and work involving documents, images and video. It has 320 billion parameters in total but activates 18 billion for each input, according to the company.

The architecture combines sparse and linear attention. Z.ai says it also reduces the number of layers compared with the similarly sized GLM-4.5 family, aiming to lower the work needed to serve requests.

Visual inputs can become feedback during a task: the model can inspect a rendered document or interface, then use what it sees to choose another action.

The announcement describes publicly available weights under the MIT license and support for SGLang, vLLM and TokenSpeed. The published performance comparisons are Z.ai’s own evaluations.