GLM-5.3-Flash: Z.ai details multimodal model with 1M-token context
Z.ai describes GLM-5.3-Flash, a natively multimodal GLM-5 model with up to one million tokens of context.
What Z.ai says about GLM-5.3-Flash
According to Z.ai, GLM-5.3-Flash has 320 billion parameters in total, with about 18 billion active for each request. Activating only part of the model helps limit computing demands. Z.ai states that the context window reaches one million tokens and positions the model for complex tasks, tool use and multistep agent workflows.
An architecture built for long context
The model combines sparse and linear attention. Sparse attention retrieves relevant information from the wider context, while linear attention helps reduce processing costs. In Z.ai’s comparison with GLM-5.3, GLM-5.3-Flash uses about three times less attention compute and 4.4 times less KV-cache memory. These are the developer’s own measurements, not an independent assessment of quality on customer tasks.
Text and visual information in one workflow
GLM-5.3-Flash was trained on text and visual information together. Z.ai says it supports images, video and files and can interpret screenshots, document layouts, charts and interface states. In an agent workflow, visual feedback can help the model inspect the result of one step before choosing the next.
Open weights and deployment
Z.ai has published the GLM-5.3-Flash model weights under the MIT License and lists SGLang, vLLM and TokenSpeed as supported inference frameworks. Check the official model repository for current compatibility and deployment terms.
How Z.ai prepared its infrastructure
Z.ai says production inference for GLM-5.3-Flash runs on a cluster of more than 100,000 Chinese-made AI accelerators. An Infra Agent powered by GLM-5.3 helped optimize the system. According to the company, moving from initial model adaptation to production readiness took under two weeks, and end-to-end serving throughput tripled against the initial baseline on the same hardware.
Why this matters to NEXSTA
GLM-5.3 and GLM-5.3-Flash are relevant to NEXSTA when assessing long context, tool use, response speed and visual inputs. This article describes capabilities reported by Z.ai. It does not announce a new multimodal feature in NEXSTA.
For more on these models and their role in the product, read GLM and Z.ai: what these models are and Models used by NEXSTA.