Last updated: November 17, 2026
Liquid AI d1.2 and d1.2-ultra are hosted decision models released November 16, 2026. They add audio input and five typed tasks: reading values, finding passages, locating objects, estimating numbers and multi-select choices. Pricing is $0.02 and $0.042 per million input tokens, with free output.
Contents
- What changed in d1.2
- d1.2 vs d1.2-ultra vs d1.1
- How to make your first d1.2 call
- The five new typed answers
- Images and audio: limits and token costs
- How d1.2 compares with other decision models
- Limitations to know
- Frequently asked questions
What changed in d1.2
A decision model does not write prose. It evaluates an input, called the state, and returns probabilities over fixed outcomes with zero generated tokens. Earlier d1 models answered three question types: Noul (yes or no), Choice (pick one) and Score (a rubric position).
According to Liquid AI’s November 16 announcement, d1.2 and d1.2-ultra keep those types and add three things:
- Five typed tasks: read a value, find every passage, locate objects, estimate a number and pick every option that applies.
- Audio input: a recording can be what you ask about, or the question itself can be spoken.
- Higher benchmark scores: d1.2-ultra scored 75 on the Artificial Analysis Decision Index v2.1, which Liquid AI says is two points ahead of the previous best decision model.
If you followed our LFM3 guide to Liquid AI’s on-device models, d1.2 is the hosted side of the same decision-model family.
d1.2 vs d1.2-ultra vs d1.1
The figures below come from the d1.2 model documentation and the deprecations page.
| Item | d1.2 | d1.2-ultra | d1.1 / d1 / d1-ultra |
|---|---|---|---|
| Status | Default hosted model | Most capable hosted model | Deprecated, still served |
| Inputs | Text, images, audio | Text, images, audio | No audio |
| Question types | Noul, Choice, Score plus five typed tasks | Noul, Choice, Score plus five typed tasks | Noul, Choice, Score only |
| Input price (per 1M tokens) | $0.02 | $0.042 | Not covered here |
| Output price | Free (zero output tokens) | Free (zero output tokens) | Zero output tokens |
| Input limit per question | 65,536 tokens | 65,536 tokens | Not covered here |
| Migration | Replaces d1 and d1.1 | Replaces d1-ultra | Change the model field |
A single call can carry up to 128 questions, and each question is also capped at 250,000 bytes of text.
How to make your first d1.2 call
The setup steps are short:
- Go to console.liquid.ai, register if needed and join an organization.
- Open Dashboard > API Keys, create a key and copy it. Keys start with
liquid_. - Store it as the environment variable
LIQUID_API_KEYrather than pasting it into code or prompts. - Send a POST request to
https://api.liquid.ai/decisions/v1/systemonewithmodelset tod1.2ord1.2-ultra, astateand aquestionsobject. - Give each question a name, a
typeandinstructions. Choice and Score questions also takecriteria. - Read the
answersobject, keyed by question name, and checkusage.input_tokens.usage.output_tokensis always 0.
In Liquid AI’s invoice example, one request asks for the payee, the total, the due date and the VAT. The model returns “ACME SA”, 1234.5, 2026-04-03 and null for VAT, because the invoice does not state it. That “not found” behavior is the main reason to use a decision model for extraction.
The five new typed answers
All five use the same endpoint and request format. Earlier models return HTTP 422 on these questions.
- Read a value (
text,number,date): returns the value, the characters it came from and theirspan. Values are read, not converted, so “1 234,50 EUR” becomes 1234.5 with unit EUR. - Find every passage (
spans): returns each matching passage with its position, useful for redacting personal data. The defaultmaxis 16 and can go up to 64. Ifat_leastis true, raisemax. - Locate objects (
box,boxes): returns box edges as fractions of the picture’s height and width, plus a count forboxes. The request must contain exactly one image. - Estimate a number (
numberwithrange): returns a median and a 10th to 90th percentile interval. Liquid AI says estimates are reliable on counts and scores, less so on free quantities. - Pick several options (
choicewithmultiple: true): scores each option on its own, so probabilities do not sum to 1. Options at 0.5 or higher are listed inchoices.
On Liquid AI’s internal test sets, d1.2-ultra read receipt values correctly 98% of the time, against 96% for d1.2. These are vendor-run results, since no public benchmark covers the new tasks yet.
Images and audio: limits and token costs
Both models bill input tokens only. A request bills its state, images and audio once, and each extra question adds its own tokens. A spans or boxes question at the default max of 16 costs about 250 and 415 tokens respectively.
- Images: JPEG, PNG or WebP, sent as Base64. Remote image URLs are not accepted. Up to 8 images per request, and up to 10,000 patches of 32 x 32 pixels in total. Each patch costs 1.5 tokens, so a 1024 x 1024 image costs 1,536 tokens.
- Audio: 16-bit PCM WAV in Base64, sent in the
mediaarray. Up to 8 clips and 300 seconds per request. Audio costs 10 tokens per second, so a 42-second call costs 420 tokens. - Request size: keep the whole JSON body under 4.5 MB. Mono 16 kHz audio fits about 100 seconds.
How d1.2 compares with other decision models
Liquid AI measured calibration error, the gap between a model’s stated probability and how often it is right, across 13 benchmarks and 22,505 questions. Lower is better.
- d1.2-ultra: 0.8 points
- d1.2: 1.2 points
- Jev 1.14: 2.7 points
- Perplexity Decider v1.3: 3.7 points
- GPT-6 Luna on OpenAI’s Decisions API: 7.2 points
These are Liquid AI’s figures. For competitors, see our guides to Perplexity Decider v1.3, Gemini Decision 1 and Hindsight decision models. If you prefer self-hosting, our vLLM 0.33 guide covers its /v1/decisions endpoint.
Liquid AI’s decision models overview recommends them for classification, routing, scoring, moderation and guardrails, and a language model for conversation, summarization and code generation.
Limitations to know
- Hosted only. There are no d1.2 weights. For local use, Liquid AI points to open-weight d1-2B, d1-omni-2B and d1-omni-600M.
- SDK gaps. The TypeSafe SDKs drop typed answers, and a multiple-choice question makes the call fail. Use direct HTTP calls for the new tasks.
- Reads come from text. A value that appears only in an image is not found. Send OCR or PDF text as the state.
- Distribution. OpenRouter and Vercel support is listed as coming soon, not available at launch.
Frequently asked questions
What is Liquid AI d1.2?
d1.2 is a hosted decision model on the Liquid API, released November 16, 2026 alongside d1.2-ultra. Instead of generating text, it returns calibrated probabilities and typed answers for questions about text, images and audio, with zero output tokens.
How much does d1.2 cost?
Liquid AI lists d1.2 at $0.02 and d1.2-ultra at $0.042 per million input tokens. Output is free. Images cost 1.5 tokens per 32×32-pixel patch and audio costs 10 tokens per second, both at the input rate.
What is the difference between d1.2 and d1.2-ultra?
Both use the same endpoint, question types and response format. d1.2-ultra is the more capable model for harder decisions and scored 75 on the Artificial Analysis Decision Index v2.1. d1.2 is faster and costs less than half as much.
Can I download d1.2 weights?
No. d1.2 and d1.2-ultra are hosted on the Liquid API. Liquid AI separately offers open-weight decision models, d1-2B, d1-omni-2B and d1-omni-600M, under the LFM Open License for local use.
Is d1.1 still available?
Yes, but it is deprecated. d1.1, d1-ultra and d1 still answer Noul, Choice and Score questions, but not the new typed questions or audio. To migrate, change the model field to d1.2 or d1.2-ultra.
Sources
- d1.2: State-of-the-Art Omnimodal Decision Models, Liquid AI Blog (November 16, 2026)
- d1.2, Liquid AI Docs
- Decision Models, Liquid AI Docs
- Deprecations, Liquid AI Docs
Featured image: Liquid AI logo, public domain (trademarked), via Wikimedia Commons. Liquid AI is a trademark of Liquid AI, Inc. FreeAnswers is not affiliated with Liquid AI.