RefCOCO-M - release-reported, metric unspecified
Multimodal
RefCOCO-M is Moondream's cleaned version of the RefCOCO (UNC) validation split, with masks re-segmented by a model ensemble, unrecoverable instances dropped and harmful expressions removed; its authors score mask IoU. Release tables that print a RefCOCO-M number without defining the metric are kept as a diagnostic; Liquid AI's values sit well above any published mask score, which suggests box precision instead.
Top models (higher is better)
| Model | Score |
|---|---|
| LFM3-VL-2B | 95.2 |
| Qwen3.5-4B | 93.3 |
| Qwen3.5-2B | 90.6 |
| GPT-5.6 Luna | 89.5 |
| Gemma 4 12B IT | 88.8 |
| Gemini 3.5 Flash-Lite | 87.6 |
| LFM2.5-VL-450M | 81.3 |
| Gemma 4 E2B IT | 76.5 |