Benchmarking Vision LLMs for Meal Calorie Estimation Using External Food Databases

A developer benchmarked several multimodal LLMs on estimating meal calories from images and text descriptions using 25 meals from the Nutrition5k dataset. Equipped with access to food composition tools, Muse Spark 1.3 performed best, achieving 48% of its calorie estimates within a 20% error margin. The test demonstrates how vision-capable LLMs paired with external database lookup tools can assist with automated nutrition tracking. It also shows that a model's suitability for specialized tasks on consumer hardware does not strictly depend on parameter size. Models evaluated calorie contents using access to USDA FoodData Central and Japan's MEXT food composition databases. DeepSeek v4 Flash Vision ranked second with 40% accuracy within the 20% error margin, while Muse Glimmer 30b significantly outperformed Qwen 3.8 27b despite similar hardware requirements.