Skip to content

Liquid AI’s d1.2 Decision Models Add Audio and Structured Answers

News
Black field audio recorder beside a vivid blue mountain lake photograph on a warm wooden desk.
Liquid AI released d1.2 and d1.2-ultra on November 16. The models can now take sound as well as text and images. They also add five ways to give answers in a set form. An app can use them to read a bill, tag a photo or route a support call.
These models are built to make one choice at a time as part of a larger job. An app can ask a yes-or-no question or give the model a list of choices. The model can return a score to show how sure it is. The app can use that score to choose what to do next. See the Liquid AI launch post.

What Liquid AI d1.2 can do

An app can ask for a date or sum shown on a bill. The model can return the field and the part of the bill it read. It can also say when the bill lacks that field. Another answer type finds each part of a text that fits a request, with a note on where it appears.
For a photo, the model can draw a box around each object that fits a request. It can also give an estimate with a range. The fifth new answer type lets it pick more than one label. A note about a late order and a wrong charge could get both a shipping and a billing tag.
Sound gives apps one more way to use these choices. A team can ask which language a clip is in. Its app can then send the clip to the right tool to turn speech into text. It can send clips with low scores to a person for review. Our Akamai coverage looks at where the model’s role ends and the app’s work begins.

How the two models scored

At launch, d1.2-ultra led the Artificial Analysis Decision Index with 75 points out of 100. The base model scored 69. It took 150 milliseconds per task on the index, while Ultra took 218. The index tests tasks such as reading a form or picking a tool. It also checks how sure the model is when it gets an answer wrong. See the full results and test criteria.
Liquid AI used its own tests to check the five new answer types. In its test of fields on receipts, Ultra got 98 percent right. The base model got 96 percent right. How well this works for other teams can depend on the forms, images or clips they use.

What access costs

Both models are live on the Liquid API. It costs $0.02 per million input tokens for d1.2 and $0.042 for Ultra. The fee is based on what an app sends in. The size of an image affects how much it costs to send. An app can ask more than one question about the same image in a single call. The service counts the image once and adds the text for each question. Liquid AI says it will add access through OpenRouter and Vercel.
loading

Loading