Text and pictures can now go into d1-ultra, a model that returns probabilities instead of a paragraph. Liquid AI released it on October 15 for decisions such as checking a circuit-board image against a specification.
The distinction changes how an application uses the result. A system that needs to sort claims or flag a possible defect can ask a fixed question and receive a score. It does not need to extract a decision from a conversational answer.
The model supports yes-or-no questions, choices among labels and positions on a scale. Liquid says it processes the input in one forward pass without generating output tokens, the small units normally produced by a language model.
The listed API price is $0.042 per million input tokens. Images count toward that input bill. Each question is charged as a separate prompt, including the associated text and images, so several questions repeat those input costs.
Liquid reports comparisons across 112 evaluations and says it trained the model to make confidence scores better match actual accuracy. These are the developer's results, not a guarantee of calibrated probabilities in an unfamiliar setting.
A probability is useful only in relation to the decision being made. A manufacturer would still need to measure missed defects and false alarms on its own production images.
The release is available through Liquid's API. Access through Vercel and OpenRouter is described as forthcoming. The published price covers input tokens only.
