Liquid AI ships d1.2 decision models with audio input and five new task types, as d1.2-ultra leads the Decision Index by two points
Synopsis
Liquid AI released the decision models d1.2 and d1.2-ultra, adding audio input and five structured-output tasks: reading a value from text, extracting spans, locating objects in images, estimating a number with a range, and selecting every applicable option. On the Artificial Analysis Decision Index v2.1, d1.2-ultra scores 75, two points above the previous best decision model; across 13 benchmarks and 22,505 questions, its calibration error is 0.8 points, versus 1.2 for d1.2 and 7.2 for GPT-6 Luna on its Decisions API. No public benchmark yet measures the five new tasks, which were evaluated on internal test sets built from held-out splits of public datasets.
Interpretation
Decision models can now take audio directly and constrain answers to checkable structured forms: a read value comes back with the characters it was read from, and a missing target returns "not found" instead of a guess. Earlier versions of the decision models handled text and image input only and did not offer the answer types these five tasks require. On internal test sets built from held-out splits of public datasets, d1.2-ultra reads a value from a receipt at 98% accuracy and d1.2 at 96%; on Nemotron PII, personal-data extraction reaches 94%/93% and 92%/88% precision/recall respectively.
Both models expose the new answer types through the same endpoint and request format, with each question declaring a type and instructions and results indexed by question name. The new answer types need no separate interface; calls follow the existing decision request format. An example request reads payee, total and due date from an invoice and returns an empty value for VAT the invoice does not state; images bill at 1.5 tokens per 32x32-pixel patch, so a 1024x1024 image costs 1,536 tokens.
Perspective
The demonstrated scope is the two models d1.2 and d1.2-ultra on the Liquid API, with text, image and audio input and five structured answer types; accuracy on the five new tasks comes from internal test sets built from held-out splits of public datasets, and the audio results are answer rates on held-out recordings with three to six options or yes-or-no questions.
No public benchmark yet measures the five new tasks, and their accuracy is reported only on internal test sets; audio results are given for three-to-six-option or yes-or-no questions and do not cover open-ended recognition. The Decision Index v2.1 lead is two points, and the cost claim is stated as no model scoring higher than d1.2 for less money.
