GLM-5: from Vibe Coding to Agentic Engineering

Summary

GLM-5 is a 744B MoE (40B active, 256 experts, 80 layers) trained on 28.5T tokens with MLA then DSA continued pretrain and 200K context. Post-train is sequential Reasoning/Agentic/General RL on slime async infra plus on-policy cross-stage distillation; 10k+ verifiable SWE envs across 9 languages. SWE-bench Verified 77.8, Terminal-Bench 2.0 56.2 (61.1 on verified), BrowseComp 62.0 / 75.9 with context management, Vending-Bench 2 $4,432; Artificial Analysis Intelligence Index v4.0 score 50 (first open-weights). Code/models https://github.com/zai-org/GLM-5.

Keywords

glm-5 · moe · dsa · agentic-rl · swe-bench · zhipu · tsinghua · coding-agents

Topics

foundation models, agentic coding, sparse attention, RL

Research notes

• Primary: arxiv abs (cs.LG; also cs.CL). CC BY 4.0 on HTML. Zhipu AI & Tsinghua. Authors field uses paper byline GLM-5 Team (HF lists Aohan Zeng / Xin Lv / Zhenyu Hou et al. plus 164 more). Code https://github.com/zai-org/GLM-5 (6,958 stars at check). HF paper page 208 upvotes; 224 linked models not copied into hf_* fields. Discord posted PDF. Training SWE/terminal environments are not a standalone public corpus on abs, so no datasets_local row. License field left blank per catalog convention.