Food2K is a large-scale food recognition dataset with 2,000 categories and over 1 million images. Compared to existing food recognition datasets, Food2K outperforms existing datasets by an order of magnitude in terms of both categories and images, thus establishing a new challenging benchmark for developing advanced models for learning visual representations of food. Food2K can be further explored to benefit more food-related tasks, including emerging and more complex tasks (e.g., nutritional understanding of food), and it can be expected that models trained on Food2K can be used as a backbone to improve the performance of more food-related tasks.