The whole dayMon 16 Nov 2026
AI Compass
WeeklyArchiveAbout

Liquid AI releases d1.2 and d1.2-ultra omnimodal decision models

Live on the 64k-context Liquid API at $0.02 and $0.042 per million input tokens with zero output fee, adding audio input and five structured extraction tasks.

Diagram showing text invoice, image and audio waveform inputs flowing into Liquid AI d1.2 and producing character span, bounding box and 0.94 refund decision outputs
Image: Liquid AI

Liquid AI launched d1.2 and d1.2-ultra on 16 November 2026 on the Liquid API (console.liquid.ai), upgrading its 64k-context decision models beyond text and images to accept spoken or ambient audio while adding five structured extraction and spatial grounding task types. Following Liquid AI’s LFM3 model family, both checkpoints bill strictly for input tokens with zero output-token charge: $0.02 per million input tokens for d1.2 and $0.042 per million input tokens for d1.2-ultra, with OpenRouter and Vercel availability planned.

Five extraction primitives, patch token metering and 0.8-point calibration

Using the existing Decisions endpoint, developers can invoke five new output types: reading exact text, number or date values with character spans (returning not found instead of guessing when absent), extracting all matching text spans such as PII, returning single or multi-instance image bounding boxes with counts, estimating quantities with numeric ranges, and multi-label option selection with per-choice probabilities. Images bill 1.5 tokens per 32x32-pixel patch (1,536 tokens for a 1024x1024 image), and multi-question requests pay for shared text and image context once.

Artificial Analysis Decision Index v2.1 bar chart ranking Liquid AI d1.2-ultra first at 75 ahead of Drex Max v1.1 at 73, Perplexity Decider v1.3 at 70, Liquid AI d1.2 at 69 and Gemini Decision 1 at 66
Liquid AI’s d1.2-ultra leads the Artificial Analysis Decision Index v2.1 at 75, while d1.2 ranks sixth at 69 ahead of Gemini Decision 1.Image: Liquid AI

On the Artificial Analysis Decision Index v2.1, d1.2-ultra placed first overall at 75 (and first on private AA-Paperclip at 76, three points clear of Drex Max v1.1), while d1.2 ranked sixth at 69, pricing 1 million Decision Index tasks at $133 and $63 respectively.

Calibration curve across 22,505 questions comparing accuracy minus confidence, showing Liquid AI d1.2-ultra at 0.8 calibration error and d1.2 at 1.2 against Jev 1.14 at 2.7, Perplexity Decider v1.3 at 3.7 and GPT-6 Luna at 7.2
Across 22,505 questions on 13 benchmarks, d1.2-ultra held calibration error to 0.8 points and d1.2 to 1.2, vs 3.7 for Decider v1.3.Image: Liquid AI

Technical details

Model IDs and context
d1.2 and d1.2-ultra on console.liquid.ai; 64k-token context window across text, image and audio
Input-only API pricing
$0.02/1M input tokens (d1.2) and $0.042/1M input tokens (d1.2-ultra); $0 output tokens
Image and multi-question metering
1.5 tokens per 32x32 patch (1,536 tokens at 1024x1024); shared context billed once per request
AA Decision Index v2.1 cost
$133 per 1M Decision Index tasks on d1.2-ultra (score 75); $63 per 1M tasks on d1.2 (score 69)

Release details & context

Held-out audio classification accuracy and extraction evaluation sets

Because public decision suites do not yet cover the five added extraction and spatial tasks, Liquid AI benchmarked them on held-out splits of CORD and SROIE receipts, Nemotron PII documents, RefCOCO/RefCOCO+/RefCOCOg visual grounding and COCO object counting, counting a span read or bounding box as correct only when it locates the target or returns not found when absent.

On held-out audio evaluations where prompts either ask about an audio clip or pose the question itself as spoken audio, both d1.2 and d1.2-ultra answered 95% to 97% of multi-way acoustic event questions on ESC-50 and 43-language spoken identification tasks on FLEURS, including close language pairs such as Spanish and Portuguese.

Common questions

How does the Liquid API bill multiple extraction questions run against the same invoice or image?

Both d1.2 and d1.2-ultra bill the shared document text and image patches once at the input token rate, with each additional question in the request adding only its own prompt tokens and zero output fee.

What do d1.2 and d1.2-ultra return when a requested field or visual object is missing from the input?

Rather than hallucinating a value or bounding box, both models explicitly return not found alongside calibrated confidence scores.

Liquid AI release timeline (3 launches tracked)

  1. Earlier13 Oct 2026
    Liquid AI releases LFM2.5-24B-A2B, an open model for agents on laptops and PCs

    Liquid AI’s LFM2.5-24B-A2B scores 27.69 on AutomationBench, against 8.72 for the larger Qwen3.6-35B-A3B.

  2. Earlier27 Oct 2026
    Liquid AI releases LFM3 open-weight model family with 256K context

    Liquid AI released the LFM3 family under open weights, providing 2B-class multimodal models with 256k context that run in 1 GB of memory.

Explore the full Liquid AI dossier or browse all model releases.

Also in the Monday 16 November edition

Read all 15 launches on one page
  1. 01
  2. 02
  3. 03
  4. 04
  5. 05
  6. 07
  7. 08
  8. 09
  9. 10
  10. 11
  11. 12
  12. 13
  13. 14
  14. 15

Spotted an error? Email aiforsuccess.in+compass@gmail.com or tag @ai_for_success on X