PaperlayerView exact revision on arXiv

Source-linked AI summary

GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models

GLM-4. 5 Team, :, Aohan Zeng, Xin Lv, Qinkai Zheng, Zhenyu Hou, Bin Chen, Chengxing Xie, Cunxiang Wang, Da Yin, Hao Zeng, Jiajie Zhang, Kedong Wang, Lucen Zhong, Mingdao Liu, Rui Lu, Shulin Cao, Xiaohan Zhang, Xuancheng Huang, Yao Wei, Yean Cheng, Yifan An, Yilin Niu, Yuanhao Wen, Yushi Bai, Zhengxiao Du, Zihan Wang, Zilin Zhu, Bohan Zhang, Bosi Wen, Bowen Wu, Bowen Xu, Can Huang, Casey Zhao, Changpeng Cai, Chao Yu, Chen Li, Chendi Ge, Chenghua Huang, Chenhui Zhang, Chenxi Xu, Chenzheng Zhu, Chuang Li, Congfeng Yin, Daoyan Lin, Dayong Yang, Dazhi Jiang, Ding Ai, Erle Zhu, Fei Wang, Gengzheng Pan, Guo Wang, Hailong Sun, Haitao Li, Haiyang Li, Haiyi Hu, Hanyu Zhang, Hao Peng, Hao Tai, Haoke Zhang, Haoran Wang, Haoyu Yang, He Liu, He Zhao, Hongwei Liu, Hongxi Yan, Huan Liu, Huilong Chen, Ji Li, Jiajing Zhao, Jiamin Ren, Jian Jiao, Jiani Zhao, Jianyang Yan, Jiaqi Wang, Jiayi Gui, Jiayue Zhao, Jie Liu, Jijie Li, Jing Li, Jing Lu, Jingsen Wang, Jingwei Yuan, Jingxuan Li, Jingzhao Du, Jinhua Du, Jinxin Liu, Junkai Zhi, Junli Gao, Ke Wang, Lekang Yang, Liang Xu, Lin Fan, Lindong Wu, Lintao Ding, Lu Wang, Man Zhang, Minghao Li, Minghuan Xu, Mingming Zhao, Mingshu Zhai, Pengfan Du, Qian Dong, Shangde Lei, Shangqing Tu, Shangtong Yang, Shaoyou Lu, Shijie Li, Shuang Li, Shuang-Li, Shuxun Yang, Sibo Yi, Tianshu Yu, Wei Tian, Weihan Wang, Wenbo Yu, Weng Lam Tam, Wenjie Liang, Wentao Liu, Xiao Wang, Xiaohan Jia, Xiaotao Gu, Xiaoying Ling, Xin Wang, Xing Fan, Xingru Pan, Xinyuan Zhang, Xinze Zhang, Xiuqing Fu, Xunkai Zhang, Yabo Xu, Yandong Wu, Yida Lu, Yidong Wang, Yilin Zhou, Yiming Pan, Ying Zhang, Yingli Wang, Yingru Li, Yinpei Su, Yipeng Geng, Yitong Zhu, Yongkun Yang, Yuhang Li, Yuhao Wu, Yujiang Li, Yunan Liu, Yunqing Wang, Yuntao Li, Yuxuan Zhang, Zezhen Liu, Zhen Yang, Zhengda Zhou, Zhongpei Qiao, Zhuoer Feng, Zhuorui Liu, Zichen Zhang, Zihan Wang, Zijun Yao, Zikang Wang, Ziqiang Liu, Ziwei Chai, Zixuan Li, Zuodong Zhao, Wenguang Chen, Jidong Zhai, Bin Xu, Minlie Huang, Hongning Wang, Juanzi Li, Yuxiao Dong, Jie Tang

arXiv:2508.06471v1cs.CL

TL;DR

A single powerful open-source model excelling across agentic, reasoning, and coding capabilities remains an unmet goal. GLM-4.5 and GLM-4.5-Air combine MoE architecture with hybrid reasoning modes, and GLM-4.5 ranks 3rd overall across 12 ARC benchmarks and 2nd on agentic tasks.

  • Problem

    A single powerful open-source model excelling across agentic, reasoning, and coding capabilities remains an unmet goal despite proprietary models’ strength in specific domains.

    Source: §1 · p. 2

  • Method

    GLM-4.5 and GLM-4.5-Air use MoE architecture and hybrid reasoning modes for complex tasks and instant responses.

    Source: §1 · p. 2

  • Results

    Across 12 ARC benchmarks, GLM-4.5 ranks 3rd overall and 2nd on agentic tasks.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The authors release GLM-4.5 and GLM-4.5-Air model weights to advance large-language-model applications and research.

    Source: §5 · p. 21

Abstract

from arXiv · show

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.

1 Introduction

The introduction presents GLM-4.5 and GLM-4.5-Air as open-source models aimed at unifying agentic, reasoning, and coding capabilities, with hybrid thinking and non-thinking modes. It positions them as addressing the lack of a single open-source model excelling across all three areas.

Motivation

  • The paper identifies agentic interaction, complex reasoning, and advanced coding as interconnected capabilities needed for real-world productivity, noting that a single open-source model excelling across all three had remained elusive.It contrasts this gap with proprietary models such as OpenAI’s o1/o3 [18] and Anthropic’s Claude Sonnet 4, which had shown strong performance in specific ARC domains.Source: §1 · p. 2

Model contributions

  • GLM-4.5 is a 355B-parameter MoE model with 32B activated parameters, and both it and GLM-4.5-Air support thinking and non-thinking modes.Thinking mode is intended for complex reasoning and agentic tasks, while non-thinking mode provides instant responses.Source: §1 · p. 2

Results

  • 70.1% on TAU-Bench, 91.0% on AIME 24, and 64.2% on SWE-bench Verified highlight GLM-4.5's agentic, reasoning, and coding performance.The introduction also reports 77.8% on BFCL v3, 79.1% on GPQA [30], 72.9% on LiveCodeBench (2407-2501), 14.4% on HLE (Humanity’s Last Exam) [28], and 37.5% on Terminal-Bench.Source: §1 · p. 2

Model contributions

  • GLM-4.5-Air has 106B parameters and matches or exceeds Qwen3-235B-A22B [47] and MiniMax-M1 [7].The introduction describes it as a significant leap among models at the 100B scale.Source: §1 · p. 2

Results

  • GLM-4.5 ranks 3rd overall across 12 ARC benchmarks and 2nd on agentic tasks, while GLM-4.5-Air ranks 6th overall.GLM-4.5 is also ranked 3rd on coding tasks and lies on the SWE-bench Verified versus model-parameters Pareto frontier; it has half the parameters of DeepSeek-R1 [13] and one-third those of Kimi K2 [34].Source: §1 · p. 2

Release

  • The models are available as open-source releases, and the authors also provide an evaluation toolkit to support benchmark reproducibility.The introduction names Z.ai, BigModel.cn, and Hugging Face as model access points and links the toolkit on GitHub.Source: §1 · p. 2

2 Pre-Training

GLM-4.5 combines a deeper MoE architecture with quality-filtered, domain-specific pre-training data and successive mid-training stages for reasoning, coding, and agentic capabilities. Training uses Muon optimization and extends context length during mid-training.

2.1 Architecture

  • The MoE design uses deeper, narrower models, loss-free balance routing, sigmoid gates, and attention choices intended to improve reasoning and support speculative decoding.It uses Grouped-Query Attention with partial RoPE, 96 heads for a 5120 hidden dimension, QK-Norm, and an MoE Multi-Token Prediction layer. The higher head count did not improve training loss but consistently improved MMLU and BBH performance.Source: §2.1 · pp. 2–3

2.2 Pre-Training Data

  • Webpage pre-training up-samples higher-quality buckets, discards the lowest-quality bucket, and uses semantic deduplication to remove similar templated pages.The highest-quality bucket contributes over 3.2 epochs; this sampling is intended to balance high-frequency reasoning knowledge and long-tail world-knowledge coverage. SemDedup [1] removes similar pages that MinHash deduplication misses.Source: §2.2 · pp. 3–4
  • Multilingual webpages and Fineweb-2 [27] documents are filtered for educational utility, with high-quality documents up-sampled.Source: §2.2 · p. 4
  • Code data is quality-tiered and sampled toward high-quality examples; all source code uses Fill-In-the-Middle, while code-related web pages undergo retrieval, quality assessment, and re-parsing.Code was curated from GitHub and other code-hosting platforms; low-quality source code samples are excluded. Retrieved web pages are re-parsed to better preserve code formats and contents.Source: §2.2 · p. 4
  • A two-stage pre-training schedule starts with general webpage documents, then up-samples GitHub code and coding-, mathematics-, and science-related webpages.Mathematics and science documents from webpages, books, and papers are scored for educational content, and documents above a threshold are up-sampled.Source: §2.2 · pp. 3–4

2.3 Mid-Training: Boost Reasoning & Agentic Capacity

  • Mid-training adds repository-level code, synthetic competition reasoning data, and long-context and synthetic agent trajectories to boost software engineering, reasoning, and agentic capabilities.Repository files and related issues, pull requests, and commits are concatenated; sequence length extends from 4K to 32K for large repositories, then to 128K with long documents up-sampled. Synthetic reasoning covers math, science, and coding competitions.Source: §2.3 · pp. 4–5

2.4 Hyper-Parameters

  • Muon is used for most parameters, with cosine learning-rate decay; training extends maximum sequence length from 4,096 in pre-training to 32,768 and 131,072 in mid-training.Muon excludes word embeddings, bias, and RMSNorm weights; its Newton-Schulz iteration count is 5, momentum is 0.95, and update RMS is scaled to 0.2. Best-fit packing is used in mid-training to avoid truncating reasoning or repository code, while pre-training uses random truncation.Source: §2.4 · p. 5, §2.3 · p. 5

3 Post-Training: Expert Model Iteration

This section describes post-training methods for building Reasoning, Agent, and General Chat experts and unifying their capabilities through self-distillation. It also presents asynchronous, decoupled RL infrastructure for long-horizon agent tasks, with continuous rollout generation, training, and a shared interface for heterogeneous agent frameworks.

3.1 Supervised Fine-Tuning

  • Removing the shortest 50% of prompts improved math and science tasks by 2%-4%; generating four responses per hard prompt added 1%-2%.SFT first cold-starts experts with extended Chain-of-Thought data, then distills millions of samples from the experts into a hybrid model with up to 128K context. The pipeline also filters samples for quality and task completion, synthesizes agentic trajectories, and uses XML-like function-call tags to reduce code escaping without harming execution performance.Source: §3.1 · pp. 6–7

3.2 Reasoning RL

  • Direct RL at the full 64K output length outperformed progressively increasing output lengths, which can cause an irreversible loss of long-context capability.The two-stage difficulty curriculum moves to extremely difficult, verified-answer problems after training plateaus on moderate problems; experiments were conducted on a smaller model. Figure 6 compares single-stage and multi-stage RL at 64K context length.Source: Figure 6 · p. 8, §3.2 · p. 8, Figure 5 · p. 8
  • Dynamic temperature increases exploration when rewards stabilize, while code and science RL benefit from token-weighted loss and expert-verified data, respectively.The temperature is set to the highest value that stays within 1% of the current validation optimum. The RL algorithm builds on GRPO without a KL loss term, and code and science ablations used a smaller experimental model.Source: §3.2 · pp. 7, 9, Figure 7 · p. 9

3.3 Agentic RL

  • Agentic RL targets web-search and code-generation agents, where automatically checkable actions and answers provide dense, reliable rewards.Web-search training data combines multi-hop knowledge-graph synthesis with human-in-the-loop extraction and selective obfuscation across web pages; software-engineering data uses GitHub issues and pull requests with executable tests in a hardened sandbox.Source: §3.3 · p. 9, §3.3.1 · p. 10

3.3.2 Pushing the Limits with Reinforcement Learning and Iterative Self-distillation

  • Agent RL uses final-answer or task-completion rewards with tool-format penalties, iteratively distills RL-generated responses into SFT data, and improves browse accuracy with more interaction turns.Only model-generated tokens are optimized, not environment feedback; the iterative process resumes RL after distillation, progressively increasing task difficulty. Figure 8 reports BrowseComp interaction-turn scaling.Source: §3.3.2 · p. 10, Figure 8 · p. 11

3.4 General RL

  • General RL combines rule-based, human, and model feedback to improve broad capabilities, instruction following, function calling, and residual output pathologies.Holistic RL uses about 5,000 prompts across 7 primary, 33 secondary, and 139 tertiary categories; instruction-following data covers 7 major and 151 minor constraint types. Function-calling RL combines exact-match step-wise rewards with end-to-end task-completion training, while targeted pathology data addresses rare errors such as language mixing and repetition.Source: §3.4 · pp. 10–12

3.5 RL Infrastructure

  • The Slime infrastructure supports synchronous colocated training for general RL and asynchronous disaggregated rollouts for agentic tasks, with FP8 inference to accelerate data generation.It supports BF16 training and online, block-wise FP8 quantization before rollout; dynamic sampling in colocated mode reduces GPU idle time, while asynchronous scheduling lets agent environments generate data continuously.Source: §3.5 · p. 13
  • Dedicated rollout and training engines run asynchronously, while a high-concurrency Docker runtime isolates tasks and avoids slow or variable trajectories blocking training.Rollout engines continuously generate trajectories; training engines update weights and periodically synchronize them back to the rollout engines, sustaining high throughput for variable agent interactions.Source: §3.5 · pp. 13–14
  • A unified HTTP interface and centralized data pool integrate heterogeneous agent frameworks by storing message-list trajectories for shared training.The pool supports task-specific filtering and dynamic sampling, while decoupling rollout logic from RL training.Source: §3.5 · p. 14
  • Together, the designs provide scalable, flexible infrastructure for long-horizon rollouts across a wide range of agent tasks.The system is designed to support long-agentic RL while adapting to diverse tasks and frameworks.Source: §3.5 · p. 14

4 Evaluation · 4.1 Evaluation of Base Models

GLM-4.5-Base’s final pre-training checkpoint is reported as stable across English, code, math, and Chinese benchmarks, supporting the authors’ approach of unifying these abilities in one model.

4.1 Evaluation of Base Models

  • GLM-4.5-Base is reported as stable across English, code, math, and Chinese benchmarks.Table 2 compares GLM-4.5-Base with representative open-source base models; the supplied passages do not provide cell values or identify a comparative winner.Source: §4.1 · p. 14, Table 2 · p. 14
  • The reported results are from GLM-4.5-Base’s last pre-training checkpoint, evaluated internally before instruction-data training.The authors say the results validate their idea of unifying all abilities into one model.Source: §4.1 · p. 14

4.2 Evaluation on 12 (ARC) Benchmarks

The section evaluates GLM-4.5 across agentic, reasoning, coding, general-ability, and safety benchmarks, reporting comparative strengths alongside evaluation scope and conditions.

4.2 Evaluation on 12 (ARC) Benchmarks

  • The full post-trained GLM-4.5 models are evaluated on 12 benchmarks spanning agentic, reasoning, and coding tasks.The benchmarks are MMLU-Pro, AIME 24, MATH-500, SciCode, GPQA, HLE, LCB (2407-2501), SWE-Bench Verified, Terminal-Bench, TAU-Bench, BFCL V3, and BrowseComp.Source: §4.2 · p. 14

4.2.1 Evaluation of Agentic Abilities

  • GLM-4.5 beats Gemini 2.5 Pro and nears Claude Sonnet 4 on TAU-bench, leads baselines on BFCL V3, and trails o3 but nears o4-mini on BrowseComp.TAU-bench covers retail and airline domains; BFCL V3 evaluates user-defined function calling, and BrowseComp tests web browsing.Source: Table 3 · p. 15, §4.2.1 · p. 15

4.2.2 Evaluation of Reasoning

  • GLM-4.5 beats OpenAI o3 on AIME 24 and SciCode, and averages above Claude Opus 4 while close to DeepSeek-R1-0528.AIME and GPQA use Avg@32 and Avg@8, respectively; HLE evaluation covers text-based questions only, with correctness judged by GPT-4o.Source: Table 4 · p. 16, §4.2.2 · p. 16

4.2.3 Evaluation of Coding

  • GLM-4.5 outperforms GPT-4.1 and Gemini-2.5-Pro on SWE-bench Verified and Claude Sonnet 4 on Terminal-Bench; on average, it is Claude Sonnet 4’s best competitor on coding tasks.Terminal-Bench evaluation uses the Terminus framework and standard function calling rather than direct prompting.Source: §4.2.3 · p. 17

4.2.4 Evaluation of General Abilities

  • GLM-4.5 matches flagship peers on MMLU, performs like 671B DeepSeek V3 and R1 on SimpleQA with 355B parameters, and beats named peers on IFEval, SysBench, and MultiChallenge.The instruction-following comparisons are against DeepSeek R1 on IFEval, GPT-4.1, DeepSeek V3, and Kimi K2 on SysBench, and GPT-4.1 and DeepSeek R1 on MultiChallenge.Source: §4.2.4 · p. 17

4.2.5 Evaluation of Safety

  • 89.87 overall, GLM-4.5 is comparable to Kimi-K2 (90.48) and GPT-4.1 (89.71), with 94.33 Ethics and Morality, 94.67 Mental Health, and 96.67 Physical Health.The passage identifies Unfairness and Bias as an area for improvement.Source: §4.2.5 · p. 17

4.3 Evaluations for Hands-on Experience

The section uses hands-on evaluation to address benchmark overfitting and assess realistic model behavior, including qualitative qualities that automated metrics may miss.

4.3.1 Evaluation of General Chat

  • The general-chat evaluation covered 660 curated real-scenario prompts in English, Chinese, and other languages, comparing GLM-4.5 with Deepseek-R1-0528 and Kimi K2.Prompts spanned seven categories; responses were randomized, and evaluators did not see GLM-4.5 or Deepseek-R1-0528 reasoning content.Source: §4.3.1 · p. 18
  • GLM-4.5 achieved the highest overall score on English prompts, the highest average on Chinese prompts, and maintained its lead in multilingual evaluation of other languages, excelling in the categories specified for each.Source: §4.3.1 · p. 18

4.3.2 Evaluation of Coding Agent

  • On CC-Bench head-to-head evaluations, GLM-4.5’s win/tie/loss rates were 40.4/9.6/50.0% versus Claude Sonnet 4, 53.9/17.3/28.8% versus Kimi K2, and 80.8/7.7/11.5% versus Qwen3-Coder.Source: Figure 12 · p. 19, §4.3.2 · p. 20
  • 90.6% tool calling success rate for GLM-4.5 exceeded Claude Sonnet 4 (89.5%), Kimi-K2 (86.2%), and Qwen3-Coder (77.1%).The evaluation also found GLM-4.5 outperformed other models in task completion consistency and agentic execution robustness.Source: Figure 13 · p. 19, §4.3.2 · p. 20

4.3.3 Evaluation of Logical Reasoning

  • Human experts found GLM-4.5 performed on par with leading models on novel logical reasoning problems.The problems were designed to require multi-step deduction and differ structurally from common online questions; experts scored response correctness and quality using a unified standard.Source: §4.3.3 · p. 20, Table 11 · p. 20

4.4 Evaluation of Translation

The section frames modern translation as requiring knowledge and reasoning about internet slang, cultural context, and domain-specific terminology, not just text conversion. GLM-4.5 is evaluated on challenging real-world cases against specialized models using blind human ratings.

  • Accurately translating “yyds” requires recognizing it as praise meaning “the eternal god,” while “胖” is a photography-community nickname for a specific Canon lens.The lens is the Canon EF 70-300mm f/4-5.6 IS USM; specialized models may mistranslate the nickname, while a general-purpose model recognizes it.Source: §4.4 · p. 20
  • A “fish” emoji can refer to the second-hand marketplace 闲鱼 (Xiányú), testing whether a model connects a visual symbol with an online cultural phenomenon.Source: §4.4 · p. 21
  • In “三花公主驾到，速来围观,” “三花” refers to calico cat coloration, yielding the idiomatic translation “The Calico Princess has arrived! Come and see!”.Source: §4.4 · p. 21
  • The blind human evaluation compared GLM-4.5 with Qwen-MT-plus, Qwen-MT-turbo, and Seed-X [9] on 100 challenging real-world cases, scoring meaning and language authenticity from 0–3.Source: §4.4 · p. 21
  • GLM-4.5 significantly outperforms specialized translation models, conveying “三花公主驾到” idiomatically where they fail contextually.The comparison is reported for selected challenging translation data; the supplied passages do not provide table values.Source: Table 12 · p. 21, §4.4 · p. 21

5 Conclusion

The GLM-4.5 series includes GLM-4.5 and GLM-4.5-Air, both using an MoE architecture that improves computational efficiency over previous GLM models. GLM-4.5 ranks third globally among open-source and proprietary models, and weights for both models are released to advance research and applications.

  • GLM-4.5 and GLM-4.5-Air both use an MoE architecture that improves computational efficiency compared with previous GLM models.Source: §5 · p. 21
  • GLM-4.5 excels at reasoning, coding, and agentic tasks and ranks third globally among open-source and proprietary models.Source: §5 · p. 21
  • The weights for GLM-4.5 and GLM-4.5-Air are released to advance large-language-model applications and research.Source: §5 · p. 21

Related papers

Loading 2508.06471v1…