glm-5

GLM-5 is a large language model designed for the demanding frontier of complex systems engineering and long-horizon agentic tasks. Built on a massively scaled architecture, it moves well beyond its predecessor with 744 billion total parameters and 40 billion active parameters during inference, trained on 28.5 trillion tokens of pre-training data. The model integrates DeepSeek Sparse Attention, a mechanism that preserves long-context understanding while meaningfully reducing the computational and deployment costs that typically come with this scale of model. The result is an architecture that can sustain extended reasoning chains and multi-step tool use without sacrificing the efficiency required for practical deployment, positioning GLM-5 as a serious option for teams building autonomous agents, coding assistants, and systems that need to operate reliably across hundreds of reasoning iterations.

The post-training pipeline leverages a custom asynchronous reinforcement learning infrastructure called slime, which was developed to address the inherent inefficiencies of scaling RL for large language models. This infrastructure substantially improves training throughput and enables more granular post-training iterations, helping bridge the gap between baseline competence and the kind of excellence required for real-world agentic workflows. The focus on iterative refinement shows up in practical use: GLM-5 sustains optimization over extended reasoning horizons and large numbers of tool calls, making it capable of tackling complex software engineering problems that demand sustained, iterative problem-solving. It delivers best-in-class performance among open-source models on reasoning, coding, and agentic benchmarks, closing the gap with frontier proprietary models. Available under an open license, it supports both commercial and non-commercial use, giving developers a powerful foundation for building agentic applications without licensing constraints.