Liquid AI’s d1-ultra makes bounded multimodal decisions without generating an answer
The API returns probabilities for fixed choices; pricing is low, while model size and independent field tests remain undisclosed.
In this article
Liquid AI has released d1-ultra, a hosted model designed for decisions with predefined outcomes rather than open-ended writing. It accepts text, images or both, and returns probabilities for bounded questions in one forward pass, with no generated output tokens. That makes the product relevant to classification, routing and scoring pipelines—but not a substitute for a general-purpose chatbot.
Three ways to ask for a decision
According to Liquid AI’s launch announcement, developers can use three question formats. “Noul” returns a probability between yes and no. “Choice” assigns probabilities to a supplied set of labels. “Score” evaluates the input against an ordered rubric. A moderation system, for example, could classify an image and caption among fixed policy categories; the model would not write a free-form explanation unless another system handled that step.
Liquid positions d1-ultra for document and image classification, request routing, triage, guardrails and model evaluation. Its decision-model documentation explicitly directs users toward a language model when the task requires conversation, summarization, code or other free-form output. The boundary matters: a probability distribution can simplify downstream automation, but it does not eliminate the need to define labels, thresholds and escalation rules.
Price and access
The launch price is $0.042 per million input tokens, with no output-token charge. Images are billed as 1.5 tokens for every 32-by-32-pixel patch; Liquid’s example prices a 1024-by-1024 image at 1,536 tokens. The October 15 announcement offered d1-ultra through Liquid’s API and said Vercel AI Gateway and OpenRouter support would follow.
Liquid describes the system as a customized, larger open-weight backbone, but it has not disclosed the exact backbone or parameter count. D1-ultra itself is API-only; unlike smaller d1 variants, its weights are not offered for download. That limits independent inspection and on-premises deployment.
What the score of 65 establishes—and what it does not
Liquid reports a score of 65 on the Artificial Analysis Decision Index and says this leads competing models by at least eight points. Artificial Analysis explains that its index combines four evaluations with equal weight, covering factual decisions and hallucination avoidance, multiple-choice reasoning, tool-use decisions and safety classification. The evaluator says it runs the tests independently.
The accessible index methodology confirms what is measured, but Liquid’s launch post is the source for d1-ultra’s numerical result. Neither the announcement nor the MarkTechPost report provides public per-test scores, and no independent production study is cited. Liquid also says d1-ultra led every category in its own 112-evaluation internal suite; that remains a vendor claim.
For buyers, the practical test is therefore specific: whether calibrated probabilities improve a real fixed-choice workflow at the chosen thresholds. The published price and interface are clear, but the undisclosed architecture, closed weights and limited public breakdown leave important questions for independent evaluation.