今天，我们很高兴地推出最新旗舰模型 GLM-4.5 和 GLM-4.5-Air，以及它们的 FP8 变体。所有模型现已上线，并在发布首日即获得 SGLang 支持。GLM-4.5 和 GLM-4.5-Air 均为功能强大的模型，旨在统一推理、编码和智能体能力，参数量分别为 355B（32B 激活）和 106B（12B 激活）。
使用 SGLang 部署 GLM-4.5
我们建议使用 SGLang 部署 GLM-4.5 系列模型以获得最佳性能。通过与 SGLang 社区的紧密合作，所有 GLM-4.5 模型从发布首日起即获得 SGLang 的全面支持。
基本用法
安装 SGLang
pip install --upgrade pip
pip install "sglang[all]>=0.4.9.post6"
355B 模型
python3 -m sglang.launch_server --model zai-org/GLM-4.5 --tp 8
106B 模型
python3 -m sglang.launch_server --model zai-org/GLM-4.5-Air --tp 8
355B FP8 量化模型
python3 -m sglang.launch_server --model zai-org/GLM-4.5-FP8 --tp 8
106B FP8 量化模型
python3 -m sglang.launch_server --model zai-org/GLM-4.5-Air-FP8 --tp 4
工具调用
在命令后附加以下参数：
--tool-call-parser glm45
推理解析器
在命令后附加以下参数：
--reasoning-parser glm45
基于 MTP 的推测解码
在命令后附加以下参数：
--speculative-algorithm EAGLE \
--speculative-num-steps [number of steps] \
--speculative-eagle-topk [top k] \
--speculative-num-draft-tokens [number of draft tokens]
GLM 4.5 模型架构与亮点
GLM-4.5 采用 MoE 架构，结合无损失均衡路由和 Sigmoid 门控，提升了计算效率。与 DeepSeek-V3 和 Kimi K2 等模型相比，我们更注重深度而非宽度——专家数量更少、隐藏维度更小，但层数更多——从而带来更优的推理性能。
关键架构设计与亮点：
- 分组查询注意力与部分 RoPE
- 针对 5120 隐藏维度设置 96 个注意力头（是常规方案的 2.5 倍），在相似训练损失下提升了 MMLU/BBH 上的推理能力
- QK-Norm 用于稳定注意力 logits
- Muon 优化器，实现更快的收敛速度和更大的批处理规模
- MTP（多 token 预测）头用于推测解码
- RL 训练由开源框架 slime 驱动，该框架由 THUDM 此前开源。
性能表现
我们在涵盖智能体（3 项）、推理（7 项）和编码（2 项）的 12 个基准测试上，将 GLM-4.5 与 OpenAI、Anthropic、Google DeepMind、xAI、阿里巴巴、月之暗面和 DeepSeek 的多种模型进行了对比。总体来看，GLM-4.5 排名第三，GLM-4.5 Air 排名第六。
智能体能力 GLM-4.5 支持 128k 上下文窗口和原生函数调用。在 -bench 和 BFCL-v3 两项基准上，它均与 Claude 4 Sonnet 持平；在 BrowseComp 网页浏览基准上，它超越了 Claude 4 Opus（26.4% 对比 18.8%），并接近 GPT o4-mini-high（28.3%）。其高达 90.6% 的工具调用成功率，凸显了它在基于智能体的工作流程中的可靠性。
推理能力 GLM-4.5 在数学和逻辑推理方面表现出色。它在 MMLU Pro（84.6）、AIME-24（91.0）和 MATH500（98.2）上取得了具有竞争力的分数，并在 GPQA、LCB 和 AA-Index 等基准测试中展现出强大的泛化能力。
编程能力 GLM-4.5 展现了全面的全栈开发能力，在 SWE-bench Verified（64.2）和 Terminal-Bench（37.5）上位列顶级模型。在直接对比评估中，它对 Kimi K2 的胜率达到 53.9%，对 Qwen3-Coder 的胜率达到 80.8%。其高智能体可靠性、多轮编程任务表现以及视觉界面质量，证明了它作为自主编程助手的强大实力。
结论
GLM-4.5 系列代表了新一代大语言模型，在长上下文推理、智能体工作流程和编程任务方面表现出色。其混合 MoE 架构——通过分组查询注意力、MTP 和 RL 训练等技术得到增强——兼具高效性和强大能力。
SGLang 提供了一个生产就绪的高性能推理栈，通过先进的内存管理和请求批处理实现无缝部署。
总之，GLM-4.5 和 SGLang 共同为下一代 AI 奠定了坚实基础——为代码、文档和智能体领域提供智能、可扩展的解决方案。
致谢
我们衷心感谢以下团队和合作者对本 PR 的贡献：
- GLM 团队：张宇轩、张晨辉、吕鑫、朱子林及各位同事。
- SGLang 团队及社区：何彪、黄立夫、江彬尧、朱明磊、万程、苏畅、童新源及众多其他贡献者。
Today, we are excited to introduce our latest flagship models GLM-4.5 and GLM-4.5-Air, along with their FP8 variants. All models are now available with day-one support on SGLang. GLM-4.5 and GLM-4.5-Air are both powerful models designed to unify reasoning, coding, and agentic capabilities, with 355B total parameters (32B active) and 106B total parameters (12B active) respectively.
Deploying GLM-4.5 with SGLang
We recommend deploying the GLM-4.5 series of models using SGLang for optimal performance. Through close collaboration with the SGLang community, all GLM-4.5 models are fully supported on SGLang starting from day one.
Basic Usage
Install SGLang
pip install --upgrade pip
pip install "sglang[all]>=0.4.9.post6"
355B Model
python3 -m sglang.launch_server --model zai-org/GLM-4.5 --tp 8
106B Model
python3 -m sglang.launch_server --model zai-org/GLM-4.5-Air --tp 8
355B FP8 Quantized Model
python3 -m sglang.launch_server --model zai-org/GLM-4.5-FP8 --tp 8
106B FP8 Quantized Model
python3 -m sglang.launch_server --model zai-org/GLM-4.5-Air-FP8 --tp 4
Tool Call
Append the following parameter to the command:
--tool-call-parser glm45
Reasoning Parser
Append the following parameter to the command:
--reasoning-parser glm45
Speculative Decoding with MTP
Append the following parameters to the command:
--speculative-algorithm EAGLE \
--speculative-num-steps [number of steps] \
--speculative-eagle-topk [top k] \
--speculative-num-draft-tokens [number of draft tokens]
GLM 4.5 Model Architecture and Highlights
GLM-4.5 adopts a MoE architecture with loss-free balance routing and sigmoid gates, enhancing compute efficiency. Compared to models like DeepSeek-V3 and Kimi K2, we prioritize depth over width—fewer experts and smaller hidden dimensions, but more layers—resulting in better reasoning performance.
Key architectural designs and highlights:
- Grouped-Query Attention with partial RoPE
- 96 attention heads for 5120 hidden size (2.5× more than typical), improving reasoning on MMLU/BBH despite similar training loss
- QK-Norm for stabilized attention logits
- Muon optimizer, enabling faster convergence and larger batch sizes
- MTP (Multi-Token Prediction) head for speculative decoding
- RL training is powered by the open-source framework slime, which was earlier open-sourced by THUDM.
Performance
We compare GLM-4.5 with various models from OpenAI, Anthropic, Google DeepMind, xAI, Alibaba, Moonshot, and DeepSeek on 12 benchmarks covering agentic (3), reasoning (7), and Coding (2). Overall, GLM-4.5 is ranked at the 3rd place and GLM-4.5 Air is ranked at the 6th.
Agentic Abilities GLM-4.5 supports 128k context and native function calling. On both -bench and BFCL-v3, it matches Claude 4 Sonnet, and on the BrowseComp web browsing benchmark, it surpasses Claude 4 Opus (26.4% vs. 18.8%) and approaches GPT o4-mini-high (28.3%). Its high tool-calling success rate (90.6%) highlights its reliability in agent-based workflows.
Reasoning GLM-4.5 excels in mathematical and logical reasoning. It scores competitively on MMLU Pro (84.6), AIME-24 (91.0), and MATH500 (98.2), and demonstrates strong generalization across benchmarks like GPQA, LCB, and AA-Index.
Coding GLM-4.5 shows comprehensive full-stack development ability and ranks among the top models on SWE-bench Verified (64.2) and Terminal-Bench (37.5). In head-to-head evaluations, it achieves a 53.9% win rate over Kimi K2 and 80.8% over Qwen3-Coder. Its high agentic reliability, multi-round coding task performance, and visual interface quality demonstrate its strength as an autonomous coding assistant.
Conclusion
The GLM-4.5 series represents a new wave of large language models, excelling in long-context reasoning, agentic workflows, and coding tasks. Its hybrid MoE architecture—enhanced by techniques like grouped-query attention, MTP, and RL training—offers both efficiency and strong capability.
SGLang provides a production-ready, high-performance inference stack, enabling seamless deployment through advanced memory management and request batching.
Together, GLM-4.5 and SGLang form a robust foundation for next-generation AI—powering intelligent, scalable solutions across code, documents, and agents.
Acknowledgement
We would like to express our heartfelt gratitude to the following teams and collaborators in this PR:
- GLM Team: Yuxuan Zhang, Chenhui Zhang, Xin Lv, Zilin Zhu and colleagues.
- SGLang Team and community: Biao He, Lifu Huang, Binyao Jiang, Minglei Zhu, Cheng Wan, Chang Su, Xinyuan Tong and many others.