Liquid AI d1-ultra: cheap decision model for text, images

What was released?

Liquid AI released d1-ultra on October 15, its new decision model after d1 (Liquid AI). It reads text, images or both. It does not write text back. It returns a probability for every possible answer.

The company built it for harder, messier real-world cases. Its two examples: spotting a defect on a circuit board against its spec, and checking a damage claim against the customer's photo.

How does a decision model work?

The input is unstructured data plus one or more questions. The model reads them in a single forward pass and returns probabilities without generating any tokens. It supports the same three question types as d1:

• yes/no questions, answered with a probability from 0 to 1;
• choice: one label from many, with a probability per label;
• score: a position on a scale.

This is not a chat model. It is meant for classification, filtering and checks where code needs a number to compare against a threshold.

Price and availability

d1-ultra is available now on the Liquid API, and you can try it in Liquid Console and the d1 Playground. It costs $0.042 per million input tokens. There are no output tokens to pay for, because the model does not generate.

Images are billed as input tokens at the same rate: 1.5 tokens per 32×32-pixel patch, so a 1024×1024 image costs 1,536 tokens. Each question is billed as its own prompt, including its text and all images. With several questions, the image is billed several times.

Liquid says the model will soon be available through Vercel and OpenRouter. It did not give a date.

What do the benchmarks show?

d1-ultra is based on a larger open-weight model than d1, with Liquid-specific customizations. Liquid says the biggest gains are on knowledge-related tasks. It compared the model with d1 and Jev 1.14 across 112 unique evaluations.

Source: Liquid AI. Results of d1-ultra and other decision models across eight task groups, measured by the vendor.

In Liquid's table, d1-ultra averages 76.7 points. Jev 1.14 scores 69.4, d3 73.5 and pplx-decider v1.1 70.3. The biggest lead is in judgment and taste (65.5 against 53.3 for Jev 1.14) and in knowledge and reasoning (56.0 against 47.8). On safety and policy the models are close: d1-ultra 85.8, d3 85.5 and pplx-decider v1.1 85.0.

On the Artificial Analysis Decision Index, the company's chart shows d1-ultra at 65 and the runner-up, Jev 1.14, at 57. Liquid says that is a lead of at least 8 points over every other decision model.

Calibration: can you trust the probabilities?

For a decision model, accuracy is not enough. The probabilities also have to match reality. Liquid says most decision models are overconfident, so it trained d1-ultra specifically to produce more reliable probabilities.

Source: Liquid AI. Accuracy minus confidence, plus mean calibration error per model.

In Liquid's chart, d1-ultra has a mean calibration error of 1.5 points. d3 is at 2.6, Jev 1.14 at 2.7, d1 at 4.6 and pplx-decider-v1.1 at 5.2.

What it means in practice

All numbers are Liquid's own measurements, with no independent check so far. The table still lists v1.1 of Perplexity's model. Liquid says d1-ultra delivers better accuracy and calibration than Jev 1.14 at the same price.

If you currently label data with a chat model and parse its reply, test d1-ultra on your own labeled sample. For image tasks, work out the cost up front, since images are billed again for every question.