Z.ai Releases GLM-5.3-Flash: A 320B-A18B Natively Multimodal MoE With a 1M-Token Context

Z.ai released GLM-5.3-Flash, a new artificial intelligence model designed to handle multiple types of input including text, images, and video. The model uses a mixture-of-experts architecture with a very large context window that allows it to process over one million tokens at once, and it performs comparably to other leading models while costing roughly one-tenth the price of its predecessor. The model matters because it makes capable AI more accessible and affordable for applications like coding, document analysis, and automation tasks, with deployment options available both through a hosted API and through self-hosting for organizations with sufficient hardware resources. The efficiency gains come from architectural innovations including a hybrid attention system that combines linear and sparse attention methods, a compression technique called IndexPool that reduces memory requirements at large context sizes, and design changes that reduce the number of active parameters needed per token.