GLM-5

GLM-5 is an open-weights model designed for long-horizon agentic engineering and complex system tasks. This 744B Mixture-of-Experts (MoE) model, with 40B active parameters, has achieved top performance among open-source models on Vending Bench 2, significantly closing the performance gap with leading proprietary models. Key features include DeepSeek Sparse Attention, which enhances efficiency and capability for processing extensive contexts, and a specialized 'slime' reinforcement learning infrastructure. These architectural choices enable GLM-5 to manage intricate dependencies and extended operational sequences crucial for advanced engineering applications. The model is built to excel in environments requiring sophisticated sequential decision-making and intricate problem-solving over prolonged periods. It offers a powerful foundation for developers and researchers working on advanced computational agents, large-scale simulations, and automated system development requiring robust and adaptable models.