We present GLM-4.1V-Thinking, GLM-4.5V, and GLM-4.6V, a family of vision-language models (VLMs) designed to advance general-purpose multimodal understanding and reasoning. In this report, we share our key findings in the development of the reasoning-centric training framework. We first develop a capable vision foundation model with significant potential through large-scale pre-training, which arguably sets the upper bound for the final performance. We then propose Reinforcement Learning with Curriculum Sampling (RLCS) to unlock the full potential of the model, leading to comprehensive capability enhancement across a diverse range of tasks, including STEM problem solving, video understanding, content recognition, coding, grounding, GUI-based agents, and long document interpretation. In a comprehensive evaluation across 42 public benchmarks, GLM-4.5V achieves state-of-the-art performance on nearly all tasks among open-source models of similar size, and demonstrates competitive or even superior results compared to closed-source models such as Gemini-2.5-Flash on challenging tasks including Coding and GUI Agents. Meanwhile, the smaller GLM-4.1V-9B-Thinking remains highly competitive-achieving superior results to the much larger Qwen2.5-VL-72B on 29 benchmarks. We open-source both GLM-4.1V-9B-Thinking and GLM-4.5V. We further introduce the GLM-4.6V series, open-source multimodal models with native tool use and a 128K context window. A brief overview is available at https://z.ai/blog/glm-4.6v. Code, models and more information are released at https://github.com/zai-org/GLM-V.
核心贡献 · Key contributions
提出 GLM-4.1V-Thinking，一种通过以推理为核心的训练框架实现通用多模态推理的视觉语言模型。 Proposes GLM-4.1V-Thinking, a VLM for general-purpose multimodal reasoning via a reasoning-centric training framework.
引入带课程采样的强化学习（RLCS），以增强跨领域推理能力。 Introduces Reinforcement Learning with Curriculum Sampling (RLCS) to enhance cross-domain reasoning capabilities.
开源 GLM-4.1V-9B-Thinking，在相似规模模型中达到最优，并与更大模型竞争。 Open-sources GLM-4.1V-9B-Thinking, achieving SOTA among models of similar size and competitive with much larger models.
展示了多领域强化学习中稳健的跨领域泛化与相互促进。 Demonstrates robust cross-domain generalization and mutual facilitation in multi-domain reinforcement learning.
开发了全面、精确的多领域强化学习奖励系统，解决了奖励欺骗和领域特定验证问题。 Develops a comprehensive, precise reward system for multi-domain RL, addressing reward hacking and domain-specific verification.
在 28 个基准测试中，10B 以下模型在 23 项上取得领先，包括 STEM、GUI 智能体和视频理解。 Achieves superior performance on 23 out of 28 benchmarks among sub-10B models, including STEM, GUI agents, and video understanding.
局限 · Limitations
强化学习提升了任务完成率，但未能持续改善推理质量；正确答案可能依赖有缺陷的推理。 RL enhances task completion but does not consistently improve reasoning quality; correct answers may rely on flawed reasoning.
强化学习训练可能不稳定；微小设置变化会导致推理深度或输出风格的显著差异。 RL training can exhibit instability; minor setup changes lead to substantial variations in reasoning depth or output style.
模型在处理杂乱图像、遮挡物体或模糊视觉细节时存在困难，导致感知错误。 The model struggles with cluttered images, occluded objects, or ambiguous visual details, leading to perceptual errors.
当前奖励模型仅评估最终结果而不评估中间推理步骤，可能强化有缺陷的推理链。 Current reward models evaluate final outcomes without assessing intermediate reasoning steps, potentially reinforcing flawed chains.
混合领域强化学习在接地和 GUI 智能体性能上的提升不如单领域训练。 Mixed-domain RL does not improve grounding or GUI-agent performance as much as single-domain training.
论文章节 · Sections（共 20）
摘要Abstract
1 引言1 Introduction
2 概述与架构2 Overview and Architecture
3 预训练3 Pre-training
3.1 预训练数据3.1 Pre-training Data
3.2 训练方案3.2 Training Recipe
4 监督微调4 Supervised Fine-Tuning
4.1 监督微调数据4.1 Supervised Fine-Tuning Data
4.2 训练方案4.2 Training Recipe
5 强化学习：挑战与有效方法5 Reinforcement Learning: What’s Challenging and What Works
5.1 数据准备5.1 Data Preparation
5.2 奖励系统5.2 Reward System
5.3 课程采样强化学习5.3 Reinforcement Learning with Curriculum Sampling (RLCS)
5.4 基础设施5.4 Infrastructure
6 评估6 Evaluation
6.1 评估设置6.1 Evaluation Setting
6.2 与其他先进多模态大语言模型对比6.2 Comparison to Other Advanced MLLMs
6.3 强化学习跨领域泛化研究6.3 Investigating Cross-Domain Generalization in Reinforcement Learning
7 讨论：局限与未来工作7 Discussion: Limitations and Future Work