# Liquid AI Releases d1.2 'Omnimodal Decision Models' With Audio Input, No Independent Benchmarks Published

> Source: <https://epyhia.com/liquid-ai-releases-d1-2-omnimodal-decision-models-with-audio-input-no>
> Published: 2026-11-17 05:08:25+00:00

# Liquid AI Releases d1.2 'Omnimodal Decision Models' With Audio Input, No Independent Benchmarks Published

Liquid AI released d1.2 and d1.2-ultra on November 16, 2026, adding audio input and five new task types to its hosted decision-model API. The launch consists of a company blog post with an invoice-processing code demo. The only independent evaluation covers the models' performance on the existing Decision Index benchmark; the new capabilities have no third-party benchmarks.

## What shipped

Both models accept text, images and audio through the Liquid API. Audio can be the material being evaluated — identifying a language or detecting a siren — or the question itself, spoken instead of typed. The API returns typed answers with probabilities rather than free-form text.

Five new answer types were added:

- **Value extraction** : read a text, number or date from the input (e.g., an invoice total), returning the source characters or "not found"
- **Passage detection** : locate every passage of a specified kind (e.g., personal data) with its position
- **Image localization** : return bounding boxes for one or all instances of an object
- **Numeric estimation** : provide an estimate with a range (e.g., 12 people, range 10–14)
- **Multi-select classification** : return every applicable label with its own probability

Both models have a 64K-token context window. Pricing is input-only: $0.02 per million tokens for d1.2 and $0.042 for d1.2-ultra, unchanged from their predecessors. Images cost 1.5 tokens per 32×32-pixel patch at the text rate. OpenRouter and Vercel integrations are said to be coming soon [\[1\]](https://thedefiant.io/frontier/ai/liquid-ai-s-d1-2-ultra-tops-decision-model-ranking)[\[5\]](https://www.liquid.ai/blog/d1-2).

## Independent benchmark: Decision Index v2.1

Artificial Analysis, which runs the Decision Index independently, reported d1.2-ultra at 75 on version 2.1 — the highest score among decision models tested. Source [\[1\]](https://thedefiant.io/frontier/ai/liquid-ai-s-d1-2-ultra-tops-decision-model-ranking) states this is four points above its predecessor d1-ultra (71). Source [\[2\]](https://hermes-ai.net/news/liquidai-s-d1-2-ultra-tops-decision-index-with-near-perfect-confidence) and Liquid's own announcement [\[5\]](https://www.liquid.ai/blog/d1-2) describe the lead as two points above the "previous best decision model" or "previous leader," suggesting a different baseline model at 73. The cheaper d1.2 scored 69, up three points from d1.1 [\[1\]](https://thedefiant.io/frontier/ai/liquid-ai-s-d1-2-ultra-tops-decision-model-ranking).

The index combines six evaluations covering document decisions and extraction, tool calling, knowledge, reasoning and consistency under rewording. Its scoring incorporates the probabilities models assign to answers, penalizing confident mistakes more than uncertain ones; a score of 75 is not a 75% accuracy rate. The private AA-Paperclip benchmark (30% weight) tests 1,500 resumes, receipts, invoices and research papers; d1.2-ultra led at 76, three points ahead of Drex Max v1.1 [\[1\]](https://thedefiant.io/frontier/ai/liquid-ai-s-d1-2-ultra-tops-decision-model-ranking).

Artificial Analysis estimated benchmark-weighted costs of $63 per million tasks for d1.2 and $133 for d1.2-ultra at current token prices [\[1\]](https://thedefiant.io/frontier/ai/liquid-ai-s-d1-2-ultra-tops-decision-model-ranking).

## New tasks: only internal tests

Liquid AI states that no public benchmark measures the five new tasks yet [\[5\]](https://www.liquid.ai/blog/d1-2). The company assembled held-out internal test sets from public datasets (CORD, SROIE, Nemotron PII, RefCOCO family, COCO) and reported the following for d1.2-ultra and d1.2 respectively [\[5\]](https://www.liquid.ai/blog/d1-2):

| Task | d1.2-ultra | d1.2 | |------|-----------|------| | Receipt value extraction (CORD, SROIE) | 98% | 96% | | Personal-data detection precision/recall (Nemotron PII) | 94%/93% | 92%/88% | | Object localization (RefCOCO family) | 94% | 94% | | Object count within one (COCO) | 93% | 92% | | Multi-label object presence (COCO) | 87% | 90% |

On audio, both models answered 95–97% of questions on ESC-50 sound identification and FLEURS language identification across 43 languages. Each question offered three to six options, including close confusions such as Spanish vs. Portuguese, or yes/no questions — narrower than open-ended audio understanding [\[2\]](https://hermes-ai.net/news/liquidai-s-d1-2-ultra-tops-decision-index-with-near-perfect-confidence).

Liquid also reports a calibration error of 0.8 points for d1.2-ultra across 13 benchmarks and 22,505 questions (all five compared models answering the same questions), versus 1.2 for d1.2, 2.7 for Jev 1.14, 3.7 for Perplexity Decider v1.3 and 7.2 for GPT-6 Luna on its Decisions API [\[5\]](https://www.liquid.ai/blog/d1-2). These are Liquid's numbers on a test it chose; no outside group has published a comparison on the same questions [\[3\]](https://cellcog.ai/blog/liquid-d1-2-decision-models).

## What remains unknown

- **No third-party evaluation of the five new tasks or audio capabilities.** The only independent scores are for the pre-existing Decision Index benchmark suite.
- **Self-hosting:** Liquid says "we will open the weights at some point" but has announced no date, license or deployment requirements[\[3\]](https://cellcog.ai/blog/liquid-d1-2-decision-models) .
- **Production characteristics:** Latency, throughput, domain shift, poor scan quality, noisy audio, and calibration on customer data are untested publicly[\[2\]](https://hermes-ai.net/news/liquidai-s-d1-2-ultra-tops-decision-index-with-near-perfect-confidence) .
- **The d1.2 vs. d1.2-ultra reversal on multi-label object presence** (90% vs. 87%) is not explained by the company[\[2\]](https://hermes-ai.net/news/liquidai-s-d1-2-ultra-tops-decision-index-with-near-perfect-confidence) .

## Context

Liquid AI's decision models compete with offerings from Perplexity, OpenAI and Jev on the Artificial Analysis index. Their combination of constrained outputs, low input prices and reported calibration below one point gives developers another option for extraction and classification pipelines that require explicit abstention and review thresholds [\[2\]](https://hermes-ai.net/news/liquidai-s-d1-2-ultra-tops-decision-index-with-near-perfect-confidence). But the new capabilities — audio input and the five answer types — ship with only the vendor's internal benchmarks.

## Sources

1. [Liquid AI's d1.2-ultra Tops Decision Model Ranking | Frontier](https://thedefiant.io/frontier/ai/liquid-ai-s-d1-2-ultra-tops-decision-model-ranking) (thedefiant.io)
2. [LiquidAI's d1.2-Ultra Tops Decision Index With Near- ...](https://hermes-ai.net/news/liquidai-s-d1-2-ultra-tops-decision-index-with-near-perfect-confidence) (hermes-ai.net)
3. [Liquid AI d1.2: Price, Benchmarks, vs Jev](https://cellcog.ai/blog/liquid-d1-2-decision-models) (cellcog.ai)
4. [LFM3-Audio-2B Context, Specs & Sources (November 2026)](https://benchlm.ai/models/lfm3-audio-2b) (benchlm.ai)
5. [d1.2: State-of-the-Art Omnimodal Decision Models | Blog | Liquid AI](https://www.liquid.ai/blog/d1-2) (liquid.ai)
6. [d1.2: State-of-the-Art Omnimodal Decision Models | Blog | Liquid AI](https://www.liquid.ai/blog/d1-2?utm_source=alphasignal) (liquid.ai)
