GLM-4 32B (0414): benchmark results
Zhipu's GLM-4-32B-0414, a 32B dense model pretrained on 15T tokens then aligned via human preference, rejection sampling and reinforcement learning, tuned for engineering code and function calling. Provider: Zhipu. Released 2025-04-14. Access: Open.
Aggregate Rating 1574 ± 11, rank #523 of 1635 rated models, from 278 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| SuiChat-CN | 74.7 | Standard F1 (self-reported) | 93 |
| EHRBench - PROMOTE | 65.85 | Accuracy (%) on the questions built from PROMOTE records, EH | 90 |
| EHRBench - Treatment | 77.9 | Accuracy (%) on the treatment-decision questions, EHR-ground | 85 |
| Evals for Every Language - Cantonese (yue) | 55.77 | Average Score (%) | 84.1 |
| EHRBench - Diagnosis | 67.09 | Accuracy (%) on the diagnosis-decision questions (complete a | 83.3 |
| SuiChat-CN - Without Context | 70.8 | Macro F1 (%) of suicide-risk level classification of Chinese | 81.4 |
| SuiChat-CN - Without Keywords | 71.6 | Macro F1 (%) of suicide-risk level classification of Chinese | 81.4 |
| EHRBench - MIMIC-IV | 66.45 | Accuracy (%) on the questions built from MIMIC-IV records, E | 76.7 |
| EHRBench - MIMIC-III | 66.27 | Accuracy (%) on the questions built from MIMIC-III records, | 73.3 |
| EHRBench - Prognosis | 53.36 | Accuracy (%) on the prognosis-decision questions (predict a | 73.3 |
| CommunityBench - Community-Consistent Generation | 233.49 | BTL-Elo rating from majority-voted LLM-judge pairwise compar | 68.8 |
| TIDE Trajectory - WebShop - AUV | 17.2 | Area Under Variation (0-100) | 68.8 |
Interactive version: theaggregate.ai/model?slug=glm-4-32b-0414 · How It Works · Data refreshed daily, snapshot 2026-10-15.