Z.ai has introduced GLM-5.3-Flash, a new multimodal model from the GLM-5 family designed for coding, AI agents and long-context workloads.
320B total parameters, 18B active parameters
According to Z.ai, the model contains 320B total parameters, with approximately 18B active during each step. GLM-5.3-Flash accepts text, image and video inputs and supports a context window of up to 1 million tokens.
The model combines sparse attention with linear attention. Z.ai also introduces IndexPool, a mechanism designed to reduce the cost of long-context processing.
Reported performance
- 63.4 versus 46.2 for GLM-5.2 on DeepSWE v1.1;
- 48.8 versus 26.2 on AutomationBench;
- 29.0 versus 29.5 for Claude Opus 4.8 on Z.ai Code Bench;
- 3× lower attention compute and 4.4× lower KV-cache size compared with GLM-5.3.
These figures were published by Z.ai and should be treated as vendor-reported results until independently reproduced.
Why it matters
GLM-5.3-Flash reflects a broader shift in AI development: model efficiency is becoming as important as raw capability. For AI agents, the cost of long sessions, document processing and multimodal inputs can determine whether a system is practical to deploy.
Z.ai says the model was served on Chinese AI chips and released under the MIT License. The weights are available through Hugging Face.
Source: Z.ai official blog.