GLM-4.5

Zhipu AI (Z.ai)'s open-weight agentic model family: GLM-4.5 (355B total / 32B active) and the smaller GLM-4.5-Air (106B total / 12B active), both Mixture-of-Experts transformers trained on 22T tokens. A hybrid reasoning model — thinking mode for complex reasoning and tool use can be switched on/off. Features function calling and a Claude-API-compatible endpoint for easy integration. 128K input / 96K output context. Official HF repo zai-org/GLM-4.5 (org rebranded from zhipuai to zai-org) ships base, hybrid-reasoning, and FP8 checkpoints under the MIT license (confirmed on the official model card). This is GLM-4.5 (2025), distinct from the newer glm-5-3 already in the catalog.

Provider-reported (HF model card): 63.2 aggregate across 12 industry benchmarks — 3rd overall among proprietary and open models, 1st among open-source. Provider-reported (company press release): SWE-bench Verified 64.2%, Terminal-Bench 37.5%, MATH 500 98.2%, AIME 2024 91.0%, GPQA 79.1%, MMLU Pro 84.6%, tool-invocation reliability 90.6% (vs Claude 89.5%, Kimi K2 86.2%). Independent: broad press corroboration of specs at launch; no independently-run benchmark replication found at research time.