Flash probably means single expert fits into some cost efficient GPU memory, and non flash models require more complex and cost expensive architectures.