A precise look at the one structural property of chat templates that multi-turn token-in / token-out depends on, an audit across major open-weights model families, and an honest accounting of the edges.
Press ? anytime to show this help