---
type: "post"
id: "liquid-ai-released-d1-2-and-d1-2-49679b"
slug: "liquid-ai-released-d1-2-and-d1-2-49679b"
title: "Liquid AI released d1.2 and d1.2-ultra, omnimodal decision models with text, image, and audio input."
description: "Liquid AI says d1.2 and d1.2-ultra add audio input and five structured decision tasks, including document extraction, image localization, and multi-label selection."
retrieval_nugget: "Liquid AI says d1.2 and d1.2-ultra add audio input and five structured decision tasks, including document extraction, image localization, and multi-label selection. The release matters to builders because it offers typed outputs and confidence-aware decisions through an API, with input-only billing and published task-specific performance claims."
published_at: "2026-11-16"
updated_at: "2026-11-16"
record_date: "2026-11-16"
date_kind: "published_at"
topics: ["models"]
entities: ["Liquid AI","d1.2","d1.2-ultra","Artificial Analysis Decision Index","OpenRouter","Vercel"]
related_patterns: ["model-answers-become-task-interfaces"]
novelty_label: "new_fact"
what_changed: "Liquid AI released d1.2 and d1.2-ultra, adding audio input and five tasks; d1.2-ultra leads Decision Index v2.1 with a score of 75."
prior_refs: ["https://newruntime.com/posts/liquid-ai-released-d1-ultra-a-multimodal-decision-3c7286/"]
source_url: "https://www.liquid.ai/blog/d1-2"
source_urls: ["https://www.liquid.ai/blog/d1-2"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:liquid-ai-released-d1-2-and-d1-2-49679b"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/liquid-ai-released-d1-2-and-d1-2-49679b/","markdown":"https://newruntime.com/posts/liquid-ai-released-d1-2-and-d1-2-49679b.md","json":"https://newruntime.com/posts/liquid-ai-released-d1-2-and-d1-2-49679b.json"}
markdown_url: "https://newruntime.com/posts/liquid-ai-released-d1-2-and-d1-2-49679b.md"
---

# Liquid AI released d1.2 and d1.2-ultra, omnimodal decision models with text, image, and audio input.

## Retrieval answer

Liquid AI says d1.2 and d1.2-ultra add audio input and five structured decision tasks, including document extraction, image localization, and multi-label selection. The release matters to builders because it offers typed outputs and confidence-aware decisions through an API, with input-only billing and published task-specific performance claims.

## Summary
Liquid AI says d1.2 and d1.2-ultra add audio input and five structured decision tasks, including document extraction, image localization, and multi-label selection. The release matters to builders because it offers typed outputs and confidence-aware decisions through an API, with input-only billing and published task-specific performance claims.

## What happened
Liquid AI announced d1.2 and d1.2-ultra as new versions of its decision models. Both are accessible through the Liquid API and add five tasks: reading a value, finding spans, locating objects, estimating counts, and selecting every applicable option. The announcement describes reading a text, number, or date from an input and returning the characters supporting the answer; if a requested value is absent, the model can return “not found.” For span extraction, it can identify each passage of a requested kind, such as personal data, and return its location in the text.

For images, the models can return a rectangle around one described object or every instance, and can report a count. They can also estimate a number with a range, or select multiple applicable choices with a probability for each. Audio is another input modality: Liquid AI gives questions about a sound or spoken language as examples, and says the question itself can be a recording rather than typed text. The new answer types use the same endpoint and request format as before. The company’s example sends multiple questions about one invoice in a single request and receives separate typed answers.

The announcement reports that d1.2-ultra scores 75 on Artificial Analysis Decision Index v2.1, two points above the previous best decision model. Liquid AI says the index covers six benchmarks, including extraction and tool calling, and scores the probability assigned to answers so confident errors count more. Across 13 benchmarks and 22,505 questions answered by five models, the company reports calibration error of 0.8 points for d1.2-ultra and 1.2 for d1.2. Its comparison lists 2.7 for Jev 1.14, 3.7 for Perplexity’s Decider v1.3, and 7.2 for GPT-6 Luna on its Decisions API.

For the five new tasks, Liquid AI says it used internal tests built from held-out splits of public datasets because no public benchmark measures them yet. Reported d1.2-ultra results include 98% for reading receipt values, 94% precision and 93% recall for finding personal data, and 94% for locating a described object. On the count task, it reports 93% within one; for selecting every object present, 87%. Both models reportedly answered 95% to 97% of audio questions on held-out recordings. These are company-reported results, not a public independent evaluation of the new tasks.

Liquid AI lists d1.2 at $0.02 and d1.2-ultra at $0.042 per million input tokens. Both bill input tokens only. An image costs 1.5 tokens per 32×32-pixel patch at the text rate, making a 1024×1024 image 1,536 tokens. The request bills text and images once, while each extra question adds only its own text. Liquid AI says the models are coming to OpenRouter and Vercel; its announcement does not give a date for that availability.

## What the archive adds
New Runtime’s earlier coverage described d1-ultra as a probability-bearing decision interface alongside generated-text models ([October 21, 2026 note](https://newruntime.com/posts/liquid-ai-released-d1-ultra-a-multimodal-decision-3c7286/)). Another note covered Liquid AI’s broader LFM3 model family, including decision models ([October 27, 2026 note](https://newruntime.com/posts/liquid-ai-launches-lfm3-an-open-weight-2b-d00d7d/)). This is a new release and a new fact on a known story: d1.2 extends the decision-model interface with audio and five structured tasks.

## Trend signal
This release supports the shift toward task-specific outputs rather than treating natural language as the only model interface. Liquid AI’s typed answers—values, spans, boxes, ranges, and multi-select choices—make outputs more directly consumable by software, while confidence scores can inform review thresholds. The connection is to [model answers becoming task interfaces](https://newruntime.com/patterns/model-answers-become-task-interfaces/); this release is a concrete product example, though its reported task accuracy still needs independent validation.

## Why now
Liquid AI frames the release around expanding decision models to audio and five additional tasks, while reporting a first-place result on the Decision Index v2.1. The announcement is dated November 2026. It does not explain a broader launch timetable for its planned third-party distribution.

## Connections to other events
Earlier New Runtime coverage described Liquid AI’s d1-ultra as a probability-bearing decision interface. The new versions preserve that direction while adding concrete structured task types and audio input. Liquid AI also says the models are expected on OpenRouter and Vercel, extending their distribution beyond its own API when those integrations arrive.

## Why it matters
Typed outputs can reduce the glue code needed to turn a model’s prose into application data. A team processing invoices, for example, can ask for several named fields in one request and receive typed values; it can also ask for a missing value and handle a “not found” result rather than treating a plausible guess as fact. Span positions, boxes, and explicit ranges similarly offer interfaces that downstream systems can validate or route.

The probability-bearing design is potentially useful for allocating human review: teams can set confidence thresholds and send uncertain cases to a reviewer. But calibration quality is task- and population-dependent, so the reported score is not a substitute for testing on a team’s own documents, languages, images, and audio. Input-only pricing and shared billing for multiple questions may make batching worth evaluating, but actual cost depends on tokenized text and image patches. The practical gain is therefore a new option to test, not evidence that these tasks can be safely automated without oversight.

## Limits and uncertainty
The performance figures come from Liquid AI’s announcement. The company says there is no public benchmark for the five new tasks and reports internal evaluations on held-out splits; independent replication is not established here. Dataset scores may not predict results on a particular team’s documents, image conditions, audio, or label definitions. The announcement gives calibration results across 13 benchmarks but does not establish that calibration transfers to every new task or deployment. It also does not specify a date for OpenRouter or Vercel availability. Builders should verify current API behavior, rate limits, and total costs in their intended workload before relying on the models in production.

## Skill to build now
Build a small evaluation set from real, permissioned examples in one workflow—such as invoices with fields that are sometimes absent. Test typed extraction, explicit “not found” handling, and confidence-based review routing separately. Record field-level accuracy and review volume, then compare those results with a text-generation baseline before expanding to images or audio.

## Reader takeaway
d1.2 turns Liquid AI’s decision-model approach into a broader set of structured, multimodal API tasks. Its most useful promise is not simply a higher index score, but the ability to request outputs your application can validate and route. Treat the reported results as a reason to run a task-specific evaluation, not as a guarantee of production accuracy.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Models](https://newruntime.com/topics/models.md): Explore the models topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)
