AI technology

Liquid AI adds audio and typed extraction to d1.2 decision models

Liquid AI's November 16 release extends its hosted decision models beyond selecting options. The practical question is whether their typed outputs and missing-value behavior fit a workflow—not whether a company benchmark makes every decision dependable.

By SpicyRanksPublished 2 linked sources ↓

Event date: Nov 16, 2026

More than choosing an option

Liquid AI announced d1.2 and d1.2-ultra on November 16 through its hosted API. The release adds audio input and five task categories, covering values, text spans, image boxes, numerical estimates and multiple applicable choices. [1]

The launch lists $0.02 and $0.042 per million input tokens respectively, with input-only billing. Its accuracy results for the newly added tasks come from internal tests using held-out portions of public datasets; they are not independent evidence of performance on a reader's workload. [1]

Missing and uncertain are different states

Current documentation distinguishes extracted values from estimates. Missing extracted values use a found threshold; a value that cannot be parsed can retain its source span while its parsed value is null. Estimated quantities instead use a range and do not have a found field. Box queries require exactly one picture. [2]

Spicy's takeAnalysis: that distinction matters when connecting an output to another system. A missing invoice value should not silently become zero, and a numerical estimate should not be treated as a copied source value. Define how each state reaches review or a fallback before using it in an automated decision. [2]

Availability without a performance promise

Liquid's documentation lists the two d1.2 models as hosted services and separately lists smaller open-weight models. It recommends language models for free-form writing and conversation rather than these structured decision tasks. This report has not tested latency, accuracy or account access. [2]

Spicy's takeAnalysis: start with the output contract you need, then evaluate it on representative and absent-answer cases. The release is useful because it offers additional input and answer types; whether those types are reliable enough for your application remains a separate evaluation. [1][2]

Sources & further reading

Source material checked Nov 21, 2026. Linked material may change after publication.

  1. d1.2: State-of-the-Art Omnimodal Decision Models ↗

    Liquid AI · Primary source · Nov 16, 2026

  2. Decision Models documentation ↗

    Liquid AI · Primary source