GLM-4.6V-Flash: benchmark results
Zhipu's GLM-4.6V-Flash, a lightweight 9B vision-language model for local, low-latency deployment with native multimodal function calling for agentic use. Provider: Zhipu. Released 2025-12-08. Access: Open.
Aggregate Rating 1599 ± 11, rank #1019 of 10916 rated models, from 284 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| FindIt - Instance Detection | 60 | Average F1@0.5 (%) over instance detection (find the specifi | 100 |
| FindIt - Referring Expression Detection | 68.3 | Average F1@0.5 (%) over referring expression detection on 1, | 100 |
| FindIt - Video Detection | 50.9 | Average F1@0.5 (%) over video detection (localize the target | 100 |
| HM-Bench (RGB) | 44.46 | Average accuracy (%) over all tasks of HM-Bench's 19,337 hyp | 100 |
| LinkS2Bench - Target Visibility | 85.9 | Target Visibility: yes/no accuracy (%) on whether a marked s | 100 |
| UniRef-UAV (Image-only) | 58.1 | Precision@(F1=1, IoU>=0.5) (%; share of the image-only (exis | 91.7 |
| RetinaGroundBench - VQA Recall | 74.9 | Recall (%) | 90.7 |
| FindIt - Single-Label Object Detection | 65.5 | Average F1@0.5 (%) over single-label object detection (all i | 88.9 |
| LinkS2Bench - Zone Counting | 71.1 | Zone Counting: mean relative accuracy (%, MRA over threshold | 88.2 |
| SIS-Bench - Action Recognition | 55.2 | Accuracy (%; 686 action recognition questions; four-option m | 88 |
| BigEarthNet.txt - Referring Point Detection | 32.16 | Mean IoU (%) of the bounding box predicted for a land-use/la | 87.5 |
| ReferFlow-Bench - Mixture - T-Acc | 10.2 | T-Acc (%) | 87 |
Interactive version: theaggregate.ai/model?slug=glm-4-6v-flash · How It Works · Data refreshed daily, snapshot 2026-11-08.