GLM-5

GLM-5 is designed to tackle complex systems engineering and long-horizon agentic tasks, representing a significant step forward in intelligence efficiency. The model architecture features a substantial scale, moving to 744B total parameters with 40B active parameters, supported by a massive pre-training dataset of 28.5T tokens. To maintain high performance while managing deployment costs, it integrates DeepSeek Sparse Attention, which preserves long-context capacity while optimizing resource usage. This design intent focuses on bridging the gap between general competence and high-level excellence in reasoning and autonomous workflows.

The development of GLM-5 leverages a novel asynchronous reinforcement learning infrastructure called slime, which improves training throughput and enables more fine-grained post-training iterations. This approach addresses the traditional inefficiencies of scaling reinforcement learning for large language models. By combining these post-training advancements with a robust pre-training foundation, the model achieves best-in-class performance among open-source alternatives in coding and reasoning. These strengths position it as a capable tool for developers requiring reliable autonomous execution and high-level problem-solving across extended task durations.