GLM-5

GLM-5 is designed to move beyond simple code generation and interface prototyping, focusing instead on building complete, functional systems and executing complex, end-to-end tasks. With a massive architecture of 744 billion parameters and 40 billion active parameters, the model is built to handle the demands of systems engineering and long-horizon agentic workflows. To maintain efficiency while supporting long-context capacity, it integrates DeepSeek Sparse Attention, which helps reduce deployment costs without sacrificing the model's ability to manage extensive information.

The model's development relies on a significant increase in pre-training data, scaling to 28.5 trillion tokens to enhance its reasoning and coding capabilities. To bridge the gap between base competence and high-level excellence, the team developed a novel asynchronous reinforcement learning infrastructure called slime. This method improves training throughput and allows for more fine-grained post-training iterations. These advancements position the model as a leader in reasoning and agentic tasks, enabling it to function as a robust tool for developers seeking to automate complex, multi-hour workflows.