Take it as is: a model (like Llama) is a Large Language Model (LLMs). LLMs can take a various sort of inputs (text, images, audio waves, whatever) and return some response. The response can be new text, an image, a suggestion of text based on what the input text was, etc. At their core, they simply take the input data, transform it into a format that they consider readable (encoding) and throw this data into an enormous washing machine that is the model itself. The model shuffles data, tries to understand relationship between data, tries to find order and patterns into what it was given as input.

Once the model is done, it presumably has found some sort of order, pattern in the data...it developed an understanding of what the provided data are and what to return as output.

This process can be helpful for a ton of tasks, particularly text processing (creation, summarisation, rephrasing). Also other form of input can be handled, but they usually require bigger washing machines (aka models) because there's more stuff to elaborate. For example, a picture is way more heavy than a text sentence, thus models that handle pictures are generally more computationally demanding (they are usually Convolutional Neural Networks, if you're interesting it that).

Models running locally are usually text processing models, because it requires relatively less power to run them.

This is what it is, at its core.