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

GLM-4.5 is an open-source Mixture-of-Experts (MoE) large language model with 355B total and 32B activated parameters, developed by Zhipu AI and Tsinghua University.

It features a hybrid reasoning method supporting both thinking and direct response modes.

Trained on 23T tokens with multi-stage pre-training and mid-training (including repo-level code, synthetic reasoning, and long-context data), it uses the Muon optimizer and loss-free balance routing.

Post-training involves expert model iteration with SFT, reasoning RL, agentic RL, and general RL, using techniques like difficulty-based curriculum learning, single-stage RL at 64K output length, and dynamic sampling temperature.

GLM-4.5 achieves strong performance on ARC benchmarks: 70.1% on TAU-Bench, 91.0% on AIME 24, and 64.2% on SWE-bench Verified, ranking 3rd overall and 2nd on agentic benchmarks among evaluated models.

A compact version, GLM-4.5-Air (106B parameters), is also released.

The models are available on Z.ai, BigModel.cn, and Hugging Face, with an open-source evaluation toolkit for reproducibility.

The paper details architecture, training data, post-training methods, and comprehensive evaluations across agentic, reasoning, coding, general, safety, and translation benchmarks, including human evaluations showing GLM-4.5 outperforming specialized translation models and strong coding agent performance on CC-Bench.

The models are released to advance research in reasoning and agentic AI systems.