GLM-4.6V: benchmark results
Zhipu's open vision-language MoE (106B total/12B active) with native tool calling and 128K context (December 2025). Provider: Zhipu. Released 2025-12-08. Access: Open.
Aggregate Rating 1648 ± 5, rank #316 of 1656 rated models, from 722 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| DefectBench (Building Facades) - External Fixings Counting Error | 0.21 | Mean absolute error of the model's count of external-fixing | 100 |
| DefectBench (Building Facades) - Material Loss Counting Error | 0.44 | Mean absolute error of the model's count of material-loss in | 100 |
| DocFormBench (Docx-Skill) | 25.57 | Format accuracy (%) on DocFormBench: share of required forma | 100 |
| IPIBench - Proactive Understanding | 27.59 | Accuracy (%) on proactive understanding (trigger on time and | 100 |
| IPIBench - Reactive-after-Proactive | 39.37 | Accuracy (%) on reactive-after-proactive interaction (answer | 100 |
| LithoBench | 94.5 | MCQ$_{\mathrm{all}}$ (self-reported) | 100 |
| MechVQA | 78.91 | Total (self-reported) | 100 |
| MechVQA - Geometric Calculation | 82.93 | Accuracy (%) on the Geometric Calculation subtask (Reasoning | 100 |
| MechVQA - Identification and Counting | 88.37 | Accuracy (%) on the Identification and Counting subtask (Rec | 100 |
| MechVQA - Projection and Multi-view | 62 | Accuracy (%) on the Projection and Multi-view subtask (Reaso | 100 |
| OmniEarth - Referring Expression Comprehension | 63 | Accuracy at IoU 0.5 (%): share of predicted boxes overlappin | 100 |
| VideoVIBE - Runtime Stability | 73.47 | Failure-diagnosis accuracy (%; questions of the Runtime Stab | 100 |
Interactive version: theaggregate.ai/model?slug=glm-4-6v · How It Works · Data refreshed daily, snapshot 2026-10-20.