Liquid AI adds vision to its d1 decision model

Liquid AI added image inputs to d1 on October 5, extending a decision model that appeared in a text preview the previous week. Developers can now use the company’s paid API to inspect a picture or make a structured decision from a screenshot. Liquid AI announcement

The model returns probabilities for predefined answers. A request can ask a yes or no question, choose among named categories, or score an input against a rubric. The same response format works for text and images, giving applications a consistent way to turn those inputs into decisions.

Access and integration limits

Image access has practical limits. The free version remains limited to text, and developers must send image requests directly because the SDK does not yet support them. The API accepts up to eight encoded images per request, with the entire request kept below 4.5 MB. Remote image links are not accepted. Liquid AI documentation

Early tests and usage costs

Liquid’s launch demonstrations include visual inspection and games played from screen images. The company reports inspection accuracy between 85 and 97 percent across four categories in the public VisA dataset. Each model received an image of a good part alongside the part being inspected. Its comparison with general language models used one run per application, so the results should be treated as an early vendor demonstration.

Billing also deserves attention. Each question is charged for all supplied images again, alongside its text. Asking several questions about one picture therefore increases the token bill even though they share the same image.

Vision is available through Liquid’s API. The Vercel and OpenRouter integrations remain limited to text for now. Liquid also says it plans to release open weights for upcoming models, but that promise does not make this a downloadable model release.

What teams should test

For teams evaluating visual classification, the useful next step is to test d1 on their own difficult examples and define when uncertain scores require review. The release makes image decisions easier to connect to software, while the application still needs a plan for wrong answers.

Illustrative photograph by Blaz Erzetic on Unsplash, published July 12, 2019, under the Unsplash License. It shows manual electronics work. No connection to Liquid AI or a d1 deployment is established. Delivered without local edits.