GLM-5

GLM-5 is designed for complex systems engineering and long-horizon agentic workflows, moving beyond conversational assistance toward autonomous software construction. Z.ai describes it as a step from "vibe coding" to agentic engineering, with the model offered through the Z.ai chat interface, the Z.ai API, and the Z.ai Coding Plan, alongside open weights on HuggingFace and a companion GitHub repository for the research community. A linked technical report on arXiv (2602.15763) provides deeper documentation of the training approach and evaluation results.

The model represents a substantial scale-up from its predecessor, expanding from 355B parameters with 32B active in the prior generation to 744B parameters with 40B active, and growing the pre-training corpus from 23T to 28.5T tokens. To keep this larger model economical at inference, Z.ai integrated DeepSeek Sparse Attention, which trims deployment cost while preserving long-context capacity for extended reasoning and codebase-scale inputs. Reinforcement learning was scaled through a custom asynchronous infrastructure called slime, enabling more fine-grained post-training iterations and contributing to marked gains on academic benchmarks relative to the earlier generation.