When Meta Platforms released the weights of its first Llama model in 2023, the act was read in different quarters as an act of philosophical generosity, a competitive disruption strategy aimed at commoditising OpenAI's primary asset, or an irresponsible proliferation of powerful technology without adequate safeguards. Three years and four major model generations later, each of these readings contains truth, and none is complete.

The open-source ecosystem that developed around the Llama architecture has produced, among its most commercially significant contributions, a set of fine-tuning techniques that allow organisations with relatively modest computational resources to achieve performance on specialised tasks that rivals proprietary models costing orders of magnitude more to produce and operate. The insurance underwriting model trained on a Llama base by a mid-sized European insurer — requiring approximately forty thousand dollars of compute and six months of specialist annotation — has, according to independent evaluations, outperformed a major commercial model on the insurer's specific documentation tasks.