Llama

Llama is a series of computer programs that can read text and write text themselves. They were developed by Meta, the corporation behind Facebook, Instagram, and WhatsApp. Such programs first learn from enormous amounts of text which words typically follow one another. Afterward, they can answer questions, summarize texts, or write program code. The name Llama originally stood for “Large Language Model Meta AI”. The first version appeared in 2023, and several successors with rising numbers have followed since.

Meta versus the closed competitors

Most well-known AI systems run exclusively on their manufacturers' servers. You send a question over the internet and get an answer back. What exactly happens inside remains the company’s secret. With Llama it’s different: Meta makes the fully trained models available for download.

This allows any company, any university, and in principle any private individual to run the model on their own hardware. For businesses, this is often the decisive point. Anyone working with patient data or contracts doesn’t want to send them to an external provider. A self-hosted Llama stays in-house.

Experts do argue, however, over whether Llama may be called “open source.” With genuine open-source software, everything is open, including the training data and usage rights. Meta only publishes the trained values of the model and attaches conditions to them. Very large platforms, for instance, are not allowed to use Llama without extra permission. The more accurate term is therefore “open weights.”

What’s actually in the file you download

A language model consists of billions of numbers, the parameters. These numbers are the result of the training and determine how the model responds. It is precisely these numbers that Meta publishes. You can think of it like the sheet music of a piece: Meta supplies the score, and your own computer produces the sound.

Llama always comes in several sizes, given in billions of parameters. Small variants with around eight billion parameters run on a good gaming PC. Large variants with seventy billion or more require special computing cards from the data center. The larger the model, the better the answers tend to be, but the more expensive it is to run.

Because the numbers are openly available, they can be modified afterward. This retraining with one’s own examples is called fine-tuning. A law firm can trim Llama toward legal language this way, a hospital toward medical reports. This is exactly how hundreds of specialized versions of Llama have come into being.

Llama in apps, chats, and stock market news

You encounter Llama most directly in Meta’s own services. The assistant in WhatsApp, Instagram, and Facebook is built on it. When a chat window there answers questions or describes images, a Llama model is usually behind it.

Much more often, Llama works invisibly in the background of other products. Numerous start-ups build their chatbots and search functions on a customized Llama version. The major cloud providers also offer Llama as a building block for programmers. Anyone wanting to test a model offline on their own laptop often reaches for a small Llama variant first.

In business news, Llama mainly appears as a strategy topic. Meta doesn’t give away the models out of idealism, but to narrow the lead of providers like OpenAI or Google. If a free alternative is good enough, prices for paid models fall. A common misconception, by the way, is that Llama is a finished chatbot like ChatGPT. It is initially just the model on the inside; everyone builds the user interface around it themselves.