Nutrient prediction on NutriBench
0UPR Energy Error (kcal)NutriMLLM (Qwen3-VL-30B)
Evaluation results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| NutriMLLM (Qwen3-VL-30B)Evaluation protocol=zero-shot, Model category=NutriMLLM (Ours), Backbone=Qwen3-VL-30B2026.06 | 0 | 0.4 | 0.4 | 0.3 | 47.4 | 48.9 | 46.3 | 54.8 | |
| BERT (+ regression head)Evaluation protocol=zero-shot, Model category=Supervised Baseline2026.06 | 0.1 | 4.5 | 0.5 | 2 | 56.6 | 77.5 | 65 | 76.1 | |
| GPT-5Model category=Proprietary MLLM Baseline2026.06 | 0.1 | 0.3 | 0.1 | 0.2 | 32.1 | 32.4 | 37.3 | 50.5 | |
| Claude Sonnet 4.5Model category=Proprietary MLLM Baseline2026.06 | 0.2 | 0.4 | 0.2 | 0.2 | 34.1 | 42.2 | 36.7 | 52.9 | |
| NutriMLLM (Qwen3-VL-8B)Evaluation protocol=zero-shot, Model category=NutriMLLM (Ours), Backbone=Qwen3-VL-8B2026.06 | 0.5 | 0.6 | 0.5 | 0.8 | 53.8 | 43.8 | 52.9 | 71 | |
| Gemini 3Model category=Proprietary MLLM Baseline2026.06 | 0.7 | 0.8 | 0.6 | 0.8 | 25.3 | 34.4 | 29.1 | 41.8 | |
| NutriMLLM (GLM-4.6V-Flash)Evaluation protocol=zero-shot, Model category=NutriMLLM (Ours), Backbone=GLM-4.6V-Flash2026.06 | 0.9 | 1 | 2.1 | 1.2 | 58.9 | 54.7 | 51.4 | 62.7 | |
| NutriMLLM (Qwen3-VL-4B)Evaluation protocol=zero-shot, Model category=NutriMLLM (Ours), Backbone=Qwen3-VL-4B2026.06 | 1.4 | 1.7 | 1.3 | 1.5 | 70.4 | 61.1 | 63 | 78 | |
| NutriMLLM (Qwen3-VL-2B)Evaluation protocol=zero-shot, Model category=NutriMLLM (Ours), Backbone=Qwen3-VL-2B2026.06 | 2.8 | 2.8 | 4.4 | 2.6 | 128.4 | 100.7 | 74.5 | 105.9 | |
| Qwen3-VL-8BModel category=Open-Weight MLLM Baseline2026.06 | 6.6 | 6.6 | 5.9 | 6.7 | 56.3 | 75.2 | 66.7 | 87.3 | |
| Qwen3-VL-4BModel category=Open-Weight MLLM Baseline2026.06 | 23.6 | 24 | 18 | 24.6 | 108.4 | 114.3 | 117.1 | 124 | |
| GLM-4.6V-FlashModel category=Open-Weight MLLM Baseline2026.06 | 25.5 | 25.2 | 25.4 | 26.8 | 75 | 82.5 | 81.2 | 95.2 | |
| Qwen3-VL-2BModel category=Open-Weight MLLM Baseline2026.06 | 39 | 39.2 | 37.5 | 39.1 | 95 | 106.1 | 97.8 | 111 |