We present GLM-4.5, an open-source Mixture-of-Experts (MoE) large language model with 355B total parameters and 32B activated parameters, featuring a hybrid reasoning method that supports both thinking and direct response modes. Through multi-stage training on 23T tokens and comprehensive post-training with expert model iteration and reinforcement learning, GLM-4.5 achieves strong performance across agentic, reasoning, and coding (ARC) tasks, scoring 70.1% on TAU-Bench, 91.0% on AIME 24, and 64.2% on SWE-bench Verified. With much fewer parameters than several competitors, GLM-4.5 ranks 3rd overall among all evaluated models and 2nd on agentic benchmarks. We release both GLM-4.5 (355B parameters) and a compact version, GLM-4.5-Air (106B parameters), to advance research in reasoning and agentic AI systems. Code, models, and more information are available at https://github.com/zai-org/GLM-4.5.
核心贡献 · Key contributions
提出 GLM-4.5，一个开源 MoE 模型，总参数量 355B、激活参数量 32B，在 ARC 任务上表现强劲。 Proposes GLM-4.5, an open-source MoE model with 355B total and 32B activated parameters, achieving strong ARC performance.
引入混合推理模式，支持思考与直接响应两种方式，实现灵活推理。 Introduces hybrid reasoning mode supporting both thinking and direct response, enabling flexible inference.
开发多阶段后训练，结合专家模型迭代与强化学习，提升智能体、推理和编码能力。 Develops multi-stage post-training with expert model iteration and reinforcement learning, boosting agentic, reasoning, and coding abilities.
在 12 个 ARC 基准上取得领先结果，综合排名第三，智能体任务排名第二。 Achieves state-of-the-art results on 12 ARC benchmarks, ranking 3rd overall and 2nd on agentic tasks.
发布 GLM-4.5 及其紧凑版本 GLM-4.5-Air（106B 参数），推动开源研究。 Releases both GLM-4.5 and a compact version GLM-4.5-Air (106B parameters) to advance open-source research.
开源评估工具包和模型，确保可复现性并方便社区使用。 Open-sources evaluation toolkit and models to ensure reproducibility and facilitate community use.
局限 · Limitations
在安全评估的公平性与偏见类别上仍有改进空间。 Performance on fairness and bias categories in safety evaluation still has room for improvement.
模型规模和计算成本可能限制在资源受限环境中的部署。 Model size and computational cost may limit deployment in resource-constrained environments.
评估主要针对英文和中文，多语言能力验证不够充分。 Evaluation focuses on English and Chinese; multilingual capabilities are less thoroughly validated.
智能体强化学习训练耗时，需要复杂基础设施支持长周期交互。 Agentic RL training is time-consuming and requires complex infrastructure for long-horizon rollouts.
混合推理模式在需要快速简洁回答的任务上可能产生次优响应。 Hybrid reasoning mode may produce suboptimal responses for tasks requiring quick, concise answers.
论文章节 · Sections（共 18）
摘要Abstract
1 引言1 Introduction
2.1 架构2.1 Architecture
2.2 预训练数据2.2 Pre-Training Data
2.3 中期训练：增强推理与智能体能力2.3 Mid-Training: Boost Reasoning & Agentic Capacity
2.4 超参数2.4 Hyper-Parameters
3 后训练：专家模型迭代3 Post-Training: Expert Model Iteration
3.1 监督微调3.1 Supervised Fine-Tuning
3.2 推理强化学习3.2 Reasoning RL
3.3 智能体强化学习3.3 Agentic RL
3.4 通用强化学习3.4 General RL
3.5 强化学习基础设施3.5 RL Infrastructure
4.1 基座模型评估4.1 Evaluation of Base Models
4.2 12 项（ARC）基准评估4.2 Evaluation on 12 (ARC) Benchmarks