Liquid AI Releases d1-ultra, a Multimodal Decision Model Scoring 65 on AA Decision Index

Liquid AI said d1-ultra is built on a larger open-weight backbone than d1 and includes heavy Liquid-specific customizations. The company did not disclose parameters or context length, and the model is available only through the hosted Liquid API, with no downloadable weights or quantized variants announced. Liquid reported that d1-ultra leads all eight categories in its 112-evaluation suite and scored 65 on the Artificial Analysis Decision Index, ranked No. 1. Those performance figures are vendor-reported, and no public per-benchmark breakdown or parameter count has been released.

A decision model answers bounded questions with probabilities instead of free text. Liquid's decision model documentation describes three question types: Noul, a yes/no question that returns a probability between 0 and 1; Choice, one label from a set with a probability per label; and Score, a probability-weighted position on an ordered rubric. The documentation recommends decision models for routing, moderation, guardrails, triage and replacing LLM-as-a-judge use cases, while recommending a language model for summarization, chat and code generation.

Each d1-ultra request carries a state, or context, optional images and one or more typed questions. The model evaluates every question against the same input in one call, and the d1 API reference shows output tokens always at zero. Liquid AI said the model was built for messy cases, giving examples such as spotting a circuit board defect against a specification and checking a damage claim against a customer's photo. The company said the gains from its decision training pipeline are largest on knowledge tasks.

Liquid AI compared d1-ultra with its earlier d1 model and TypeSafe's Jev 1.14 across 112 evaluations. According to Liquid AI's announcement, d1-ultra outperforms Jev 1.14 across all tasks at the same price. On the independent Artificial Analysis Decision Index, Liquid reports d1-ultra ranks No. 1 with a score of 65. The index averages four evaluations at 25 percent each: AA-Omniscience-Decision for knowledge and hallucination, HLE-MultiChoice for reasoning, When2Call for tool calling, and WildGuardTest for safety classification. Artificial Analysis's Oct. 13 launch article listed Jev 1.13 first at 58, with GLiDE and Microsoft-Decision-1 following at 54 each, according to MarkTechPost.

The report also compared the models' disclosed traits. Parameters are not disclosed for d1-ultra, d1, Jev, Microsoft-Decision-1 or GLiDE. d1-ultra and d1 accept text and images, while Jev accepts text and Microsoft-Decision-1 and GLiDE have undisclosed input modalities. d1-ultra and d1 are API-only, Jev is closed, and the weights for Microsoft-Decision-1 and GLiDE are not disclosed. On the Artificial Analysis Decision Index, Liquid reports 65 for d1-ultra, Jev 1.13 had 58 on Oct. 13, and Microsoft-Decision-1 and GLiDE each had 54. d1-ultra and Jev are priced at $0.042 per million input tokens, while Artificial Analysis listed Microsoft-Decision-1 at $31 per million tasks and GLiDE at $119 per million tasks. The report said d1-ultra, d1 and Jev do not bill output tokens.

Developers can use d1-ultra on the Liquid API through the decisions/v1/systemone endpoint with the model name d1-ultra. Liquid said Vercel AI Gateway and OpenRouter support are coming soon, and users can test the model first in the d1 Playground. Billing covers input tokens only. Images cost 1.5 tokens per 32-by-32-pixel patch, so a 1024-by-1024 image costs 1,536 tokens. Each question is billed as its own prompt, including its text and all images, meaning three questions on one image bill that image three times. Endpoint documentation caps requests at eight images and 10,000 image patches, and the full JSON body must stay under 4.5 MB. Liquid also claims the lowest calibration error in its tests, which it says is useful for routing thresholds.

The model's size, backbone and weights remain undisclosed, and access remains API-only, according to MarkTechPost.