Liquid AI Releases d1-ultra: A Multimodal Decision Model That Scores 65 on the AA Decision Index

Why it matters

Liquid AI has launched d1-ultra, described by the company as its strongest decision model so far. Unlike a standard language model, d1-ultra answers bounded questions with probabilities rather than free text, handling yes-or-no questions, label choices, and rubric-based scoring. It accepts text, images, or both, and produces no output tokens, which means developers pay only for input. Liquid positions it for classification, routing, moderation, triage, and as a possible replacement for judge-style language model calls. The model is built on a larger open-weight backbone than its predecessor d1, with Liquid-specific customizations, though parameters, weights, and context length remain undisclosed. Access is limited to the Liquid API, with support from other gateways promised later. Liquid reports that d1-ultra tops the Artificial Analysis Decision Index, ahead of competing decision models, and claims the lowest calibration error in its own tests. Readers should note that these performance claims are vendor-reported, with no public per-benchmark breakdown yet.