# Frontier Models A frontier model is one at the leading edge of capability at a given moment: the largest training runs, the best benchmark results, the newest behaviours. The term is relative. Last year's frontier is this year's open-weight baseline. In practice "frontier" describes a small set of labs rather than a technology. Only a handful of organisations can fund training runs at the scale that sets the edge, which is why the frontier has been mostly closed and mostly American ([[Large Language Model - LLMs]], [[Foundational Models MOC]]). ## Why the term matters for the chessboard - **The frontier is a moving target, not a moat.** Capability that is frontier today leaks outward as model output, gets [[API Distillation|distilled]] into cheaper models, and shows up in [[Open Weights]] within months. - **Frontier labs earn on the API layer.** Their revenue comes from selling tokens of the best model. When a near-frontier model is free to download, that layer gets squeezed ([[Token Economics of Inference]]). - **Frontier is defined by pretraining compute.** Much of what follows it (post-training, distillation, fine-tuning) is far cheaper. See [[Pretraining, Post-Training and Fine-Tuning]]. > [!note] Working definition > Frontier = the capability only a few can train. Near-frontier = the capability many can copy. Related: [[Chinese Frontier Labs]], [[The Cross-Pacific AI Chessboard]]