Llama

Llama is Meta’s family of language models, and its importance is less about any single version than about what it started. When Meta published the weights, the actual numbers inside the model, anyone could download a capable system and run it on their own hardware. That turned open-weight models from a research curiosity into a real alternative, and a whole ecosystem of tools for running models locally grew up around Llama specifically.

The smaller sizes, especially the 8 billion parameter versions, became the default benchmark for anything claiming to run fast or cheaply, which is why you still see them in chip announcements years later. Worth knowing: Llama is open-weight, not open-source in the strict sense. The licence carries conditions, and the training data has never been published.