ArchitectureTrainingReinforcement LearningAgents
GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models
Z.ai (Zhipu AI)·August 8, 2025
GLM-4.5 Team
View on arXivTL;DR
Open-weight MoE models (355B/32B active and a 106B/12B Air variant) with a hybrid thinking/direct mode, trained on ~23T tokens plus agentic RL, co-optimizing agentic, reasoning, and coding ability.
Why it matters
A leading open agent-native model, notable for explicitly co-optimizing the “ARC” triad in one model and open-sourcing its RL infrastructure.