Liquid AI says its d1 decision model now accepts images as well as text, letting developers ask defined questions about a picture, such as whether a circuit board is defective. Image input is available through its paid API; its free d1 service remains text-only.
A developer supplies d1 with an image and a question whose possible answers are defined in advance. The model returns probabilities for a yes-or-no answer, a label from a list, or a score on a scale, rather than composing a written response.
Liquid demonstrates the vision feature on visual-inspection tasks using the public VisA dataset. It says d1 sorted good and defective examples from four product lines—circuit boards, candles, cashews and chewing gum—with 85% to 97% accuracy. The company says the model was not trained specifically for those sorting tasks; each inspected item is shown alongside a good example from the same line.
These are company-reported demonstrations, not an industrial trial. Liquid says each of its six application tests was run once per model, and the announcement provides no independent replication or on-site results.
Sources
- Liquid AI: Introducing d1, now with vision (Oct. 5, 2026).
- Liquid AI decision-model API documentation.
- Illustrative image: Printed Circuit Board 9990.jpg by Ashley Pomeroy, licensed under CC BY 4.0. The photo depicts a sampler board, not Liquid AI's demo.
