GLM-5 is a frontier language model from Z.ai (formerly Zhipu) engineered to handle complex systems engineering and long-horizon agentic workloads. It scales its mixture-of-experts architecture to 40B active out of 744B total parameters, more than doubling the parameter count of its predecessor GLM-4.5 (355B total, 32B active), and was trained on 28.5 trillion tokens of pre-training data, up from 23T in the prior generation. The model integrates DeepSeek Sparse Attention (DSA), a mechanism designed to significantly cut deployment costs while preserving the ability to process very long contexts, making it practical for extended agentic workflows that would otherwise be economically prohibitive.
GLM-5 targets workflows where large language models function less as code assistants and more as autonomous engineers executing multi-step, end-to-end tasks. Independent benchmark coverage shows the model achieving 97.4% on MATH-500 but only 10.4% on Humanity's Last Exam, indicating strong performance on structured reasoning while revealing meaningful headroom on frontier expert-level problems. The model is distributed through Z.ai's Coding Plan subscription, with an accompanying technical report on arXiv and reference implementations available on GitHub and HuggingFace under the zai-org namespace, giving developers direct access to weights and documentation for building agentic coding systems.
