Most classification work still flows through a chatbot model that writes a sentence nobody needs. Liquid AI’s new d1 is built to skip that step entirely.
Hand it context and a set of typed questions, and it returns calibrated probabilities across a fixed list of outcomes in one call. The published output-token count is zero, because d1 never composes prose. Its answers are floats.
That fits a long list of production jobs: ticket routing, scoring, moderation, reranking and automated judging. Liquid’s own guidance is blunt. Pick a decision model when the answer is one of a known set of options, and keep a language model when the output has to be invented.
d1 runs on the Liquid API as d1:free and is API-only, so there are no weights to self-host. A request carries three parts: the model, a state that can be plain text or JSON, and the questions. Several question types can share one round trip, all judged against the same state.
Decision models do not replace language models. They claim a slice of traffic where a probability is the right answer, and they strip out the cost of writing text that only gets parsed.
