DiningBench - Protein Estimation: leaderboard

DiningBench (2026) evaluates vision-language models on multi-view user photos of restaurant dishes, testing dish classification against same-menu distractors, nutrition estimation and food visual question answering. This board is the mean absolute percentage error of the estimated protein; lower is better.

Metric: Mean absolute percentage error (%) of the estimated protein over the 1,650 DiningBench nutrition-estimation dishes (multi-view restaurant photos; ground truth from merchant labels or ingredient-based estimates cross-checked against USDA FoodData Central); lower is better. Source: arxiv.org. Saturation forecast: Around May 2027. 29 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 3 Pro (Preview)23.88#119
2Gemini 3 Flash (Preview)25.23#152
3Qwen 3 VL 30B A3B Instruct32.01#895
4GPT-4.132.42#505
5Qwen 2.5 VL 32B Instruct32.73#1062
6O4 Mini32.96#368
7Claude Sonnet 4.533.15#261
8Qwen 2.5 VL 72B Instruct33.38#841
9GPT-533.4#178
10Qwen 3 VL 8B Instruct33.87#1029
11Qwen 2.5 VL 7B Instruct34.67#1902
12GPT-4o Mini35.08#1667
13InternVL3.5-38B35.59#1021
14Qwen 3 VL 4B Instruct36#1332
15GPT-4o36.03#751

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=diningbench-protein-estimation · How It Works · Data refreshed daily, snapshot 2026-11-21.