Llama is Meta's family of foundation models, and its open-weight releases have done more than almost any other single decision to shape the current open-source LLM ecosystem: a huge share of fine-tuned, specialized, and locally-run models in use today trace back to a Llama checkpoint as their starting point, including the now-archived Code Llama for programming tasks specifically.
Meta's licensing terms are more permissive than fully closed models but carry some restrictions depending on scale and use case, a middle ground between fully open licenses like Apache 2.0 and the completely closed weights of GPT or Claude. That positioning, open enough for broad adoption and fine-tuning but not unconditionally open, has been a point of ongoing debate in the open-source AI community.
Being free to download and run, with sizes spanning small models that fit on consumer hardware up to much larger ones needing serious infrastructure, Llama remains a default starting point for teams and researchers who want to fine-tune or deploy their own model rather than build entirely from scratch or rely purely on a closed API.