Liquid AI has released d1, a decision model designed for structured choices rather than text generation. It accepts context and typed questions, returning calibrated probabilities across fixed outcomes in a single call with zero generated output tokens.
- The model offers three primitives: Noul (yes/no probability), Choice (picking from named sets with confidence), and Score (rating on an ordered rubric).
- d1 is deployable today as a hosted API under the name d1:free, but it is not trainable and has no self-hosted weights.
- It targets tasks like classification, ticket routing, scoring, moderation, reranking, and LLM-as-judge checks.
- Benefits over general LLMs include no billed output tokens, predictable latency, no schema errors, usable uncertainty via calibrated probabilities, and fewer round trips.
The model is intended to replace general LLMs for bounded tasks where the answer is one of N known options, allowing teams to keep LLMs for generation and complex reasoning.