What is a decision model?

Liquid AI's new family does not generate text. Instead, it assigns a probability to each of a fixed set of answer options for a given question. According to the company, this means no output tokens are produced, so only input is billed.

New capabilities

d1.2 adds five new task types on the same endpoint: reading a value from a document, finding every matching span, locating objects, estimating a number with a range, and picking every applicable option. Audio input is accepted as 16-bit PCM WAV, counted at 10 tokens per second.

Performance claims

Liquid AI says d1.2-ultra scores 75 on the AA Decision Index v2.1, which combines six evaluations, placing it first. d1.2 scores 69. The company reports a 0.8-point calibration error for d1.2-ultra across 13 benchmarks, meaning cases the model labels 80% likely should be right about 80% of the time. These figures come from Liquid AI's own tests.

Access and limits

The models are available only through the hosted Liquid API for now; parameter counts are not disclosed and weights are not shared. Liquid AI says it will open the weights "at some point". Limits are 65,536 input tokens per question and 128 questions per call. The earlier d1, d1.1 and d1-ultra versions are deprecated.