GLM-5
744B-total (40B active) MIT-licensed MoE with DeepSeek Sparse Attention, trained on 28.5T tokens; 77.8% SWE-bench Verified and top open model on Artificial Analysis at launch.
Doubles GLM-4.5's size and adds DSA plus an asynchronous RL stack (slime) for long-horizon agent training. Optimized for Huawei, Moore Threads and Cambricon chips for inference. Launch came a month after Zhipu's HKEX IPO (2026-01-08).
- Date
- Wednesday, 11 February 2026
- Lab
- Zhipu AI / Z.ai
- Kind
- open-weights
- Access
- open weights
- Price
- $1.00 input / $3.20 output per M tokens, 2026-02
Figures
| Measure | Value | Measured by |
|---|---|---|
| Total / active parameters | 744B / 40B 256 experts, 8 active | company |
| Pretraining tokens | 28.5T up from 23T | company |
| SWE-bench Verified | 77.8 | company |
| Humanity's Last Exam (with tools) | 50.4 | company |
| Artificial Analysis Intelligence Index v4.0 | 50 top open-weight at launch | independent |
Release 2026-02-11 (HF repo created 2026-02-11; Caixin 2026-02-12; paper v1 2026-02-17). A circulating claim that GLM-5 was trained entirely on Huawei Ascend is not supported by the sources I opened, which only describe inference adaptation to Chinese chips. Price $1.00 / $3.20 per M tokens per NYU RITS summary.
Sources
- arxiv.org/abs/2602.15763
- huggingface.co/zai-org/GLM-5
- docs.z.ai/guides/llm/glm-5
- www.caixinglobal.com/2026-02-12/zhipu-ai-launches-new-model-better-at-coding-learning-1024
- rits.shanghai.nyu.edu/ai/glm-5-zhipu-ai-ships-a-744b-open-weight-frontier-model/
This record was checked against its sources on 6 October 2026. How we check