Liquid AI launches d1.2 and d1.2-ultra, omni-modal decision models leading the AA Decision Index

Liquid AI launches d1.2 and d1.2-ultra, omni-modal decision models leading the AA Decision Index
2 min read

Liquid AI has announced the release of d1.2 and d1.2-ultra, two decision models that now accept text, images, and audio as input. According to the company’s launch post, d1.2-ultra scores 75 on the Artificial Analysis Decision Index v2.1, the highest score among decision models on that index.

A decision model does not write text. It returns calibrated probabilities over a fixed set of possible answers. Proper calibration means a model showing 80% confidence is correct about 80% of the time, allowing uncertain cases to be routed to a human. Each question is processed in a single pass, and the model generates no output tokens, so only the input is billed.

Five new answer types

The two models add five typed answer types, accessible via the same endpoint and request format. The first allows you to read a value (text, number, or date) with the characters it originates from, and answers ‘not found’ rather than guessing. The second, find each occurrence, returns each matching passage, like names or email addresses, with its position. The third is used to locate objects as bounding boxes, while the fourth estimates a number with a range, e.g., 12 between 10 and 14. The last allows you to choose all applicable options, each with its own probability.

For audio, the interface accepts WAV in 16-bit PCM and counts 10 tokens per second. On ESC-50 and FLEURS (43 languages) test recordings, both models answer 95% to 97% of questions correctly, according to Liquid AI. Per-call limits are 128 questions, 65,536 input tokens per question, and a 4.5 MB request body. The number of parameters is not disclosed, and the models are only accessible via Liquid’s hosted API. The weights are not yet published: the company states it will open them ‘at some point’.

Score and calibration

The Artificial Analysis Decision Index v2.1 combines six evaluations, including AA-Paperclip, weighted at 30%, and When2Call, weighted at 10%. On the chart shared by Liquid AI, d1.2 scores 69, two points less than d1-ultra at 71, a now-deprecated model. Liquid AI believes no model outperforms d1.2 at a lower cost.

Across 13 benchmarks and 22,505 questions, the company reports calibration errors of 0.8 points for d1.2-ultra, 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 (Decisions API). These figures come from a test conducted by Liquid AI itself, not an independent evaluation.

Pricing and deployment

Pricing is $0.02 per million input tokens for d1.2 and $0.042 for d1.2-ultra, making the former more than half the cost. New organizations receive $2 in free credit on Liquid Console. A 1024 × 1024 pixel image represents 1,536 tokens, and shared state and media are billed only once per request. The specified endpoint is https://api.liquid.ai/decisions/v1/systemone.

S’abonner
Notification pour
guest

0 Commentaires
Le plus ancien
Le plus récent Le plus populaire
0
Nous aimerions avoir votre avis, veuillez laisser un commentaire.x