Liquid AI launches d1-ultra, a model that responds with probabilities rather than text

Liquid AI introduces d1-ultra, a multimodal decision model capable of analyzing text and images, with a score of 65 on the Decision Index.

Does a product photograph match its description? Does an electronic component have a defect? Is an insurance claim consistent with the provided images? These questions do not always require a written response, but rather a decision accompanied by a confidence level. With d1-ultra, Liquid AI is developing this approach through a multimodal model capable of examining text, images, or both simultaneously. The system returns probabilities rather than sentences and ranks at the top of Artificial Analysis's Decision Index with a score of 65. Probabilities Instead of Generated Responses

d1-ultra belongs to a category of models specialized in decision-making.

Unlike conversational language models, which progressively build a text response, it examines the provided information and directly assigns probabilities to the possible answers.

Processing takes place in a single compute pass, without the successive generation of output tokens.

Liquid AI offers three types of questions.

Noul corresponds to a binary decision. The model answers a closed question with a probability between 0 and 1, for example, to determine whether a photograph shows a damaged product.

Choice allows selecting from multiple categories. The system assigns a probability to each, in order to classify a situation according to the defined options.

Score positions an observation on a scale. This feature can be used to evaluate the severity of an incident, the relevance of content, or the level of compliance of an item.

All three formats are accessible from the same API. Multiple questions can be submitted from a common context, with structured results that are directly usable in a program. Special Attention Paid to Probability Reliability

The accuracy of a decision does not depend solely on the number of correct answers obtained. To automate certain operations, it is also necessary to know whether the confidence level indicated by the model corresponds to its actual performance.

This property is called calibration.

A properly calibrated system that assigns a 90% confidence to a set of decisions should be right in approximately 90% of the cases concerned.

A model can achieve a good overall score while regularly being overly confident in its answers. This overestimation becomes problematic when a decision leads to an automated operation.

Liquid AI indicates that it specifically trained d1-ultra to reduce this gap between claimed confidence and observed accuracy.

According to its internal evaluations, the model exhibits the lowest calibration error among the compared systems.

This characteristic makes it possible to consider workflows combining multiple levels of processing. For example, a company could automate certain decisions that exceed a defined confidence threshold and forward the rest to a person responsible for reviewing them.

However, such a threshold does not guarantee the reliability of every individual decision. It must be evaluated against the data and situations specific to the given application. Top of the Artificial Analysis Decision Index

d1-ultra reaches 65 on the Artificial Analysis Decision Index, a benchmark specializing in models that produce structured decisions.

The detailed ranking gives it a score of 65.2, ahead of other models listed in this category at the time of its release.

Liquid AI also reports a lead of at least eight points over its competitors on this index.

This evaluation does not measure the same capabilities as general-purpose conversational model benchmarks. In particular, it examines tasks such as classification, response selection, recognizing situations that require action, and detecting problematic content.

d1-ultra's top spot therefore relates to a specific type of processing, rather than the entire range of capabilities typically associated with large language models. 112 Benchmarks to Measure Complex Decisions

Liquid AI also compared d1-ultra to its previous model, d1, and to Jev 1.14 across 112 benchmarks divided into eight categories.

The company reports progress in each of these categories.

The most significant differences involve tasks requiring specialized knowledge, response comparison, tool use, and decisions involving autonomous agents.

The model also supports 44 languages, enabling it to process data and instructions from different linguistic environments.

To develop d1-ultra, Liquid AI states that it started from an open-weights model larger than the one used for d1, applying significant modifications and specialized training.

The lab does not specify the name of the base model or the parameter count of d1-ultra.

The 112 benchmarks originate from Liquid AI and do not constitute a protocol fully replicated by an external organization. They should be distinguished from the Decision Index published by Artificial Analysis. Inspecting Products, Damages, and Documents

Liquid AI presents several demonstrations showcasing situations where information is incomplete or difficult to interpret.

In manufacturing, d1-ultra can compare a photograph of an electronic circuit with its specifications to identify a potential defect.

In insurance, the model can examine images submitted by a customer and assess whether they match the description of a reported claim.

Another demonstration involves e-commerce. The system compares product photographs with the information on its product page and flags possible inconsistencies.

These checks can be integrated into quality control, claims management, or content moderation workflows.

Liquid AI also illustrates its capabilities with a game of Pokémon, in which the model assigns probabilities to available actions to help select a strategy.

These examples can be viewed in the d1 Playground. They showcase usage scenarios without individually establishing the model's performance in all professional environments. An API Billed Solely on Input Data

d1-ultra is available on the Liquid API at a rate of $0.042 per million input tokens.

No output tokens are billed since the model does not generate a