zai-org/GLM-5

GLM-5 is an open-weights release from Z.AI built for complex systems engineering and long-horizon agentic tasks, where it aims to move pre-trained competence toward reliable execution on multi-step jobs. Compared with GLM-4.5, the model scales from 355B parameters with 32B active to 744B parameters with 40B active, and pre-training data is expanded from 23T to 28.5T tokens. To keep that scale deployable, GLM-5 integrates DeepSeek Sparse Attention (DSA), a sparse-attention design that reduces inference cost while preserving the ability to handle long inputs.

On the post-training side, Z.AI developed slime, an asynchronous reinforcement-learning infrastructure intended to make large-scale RL iterations more efficient and fine-grained. With this combination of broader pre-training and more capable RL, GLM-5 is reported to improve substantially over GLM-4.7 on a wide range of academic benchmarks, and to reach best-in-class results among open-source models on reasoning, coding, and agentic tasks, narrowing the gap with closed frontier systems. That emphasis on long-context reasoning and tool-driven workflows makes GLM-5 a practical fit for engineering assistants, multi-step automation, and other agentic pipelines where sustained planning matters more than single-turn fluency.