Liquid AI Releases d1-ultra: Multimodal Decision Model Scores 65 on AA Index

Liquid AI has launched a rather novel model — d1-ultra. Unlike today's mainstream conversational generation models, it positions itself as a "decision model," handling both text and image inputs and directly outputting calibrated probability values in a single forward pass. According to MarkTechPost, the model scores 65 on the Artificial Analysis Decision Index, with an input cost of $0.042 per million tokens.

What It Means to Output Calibrated Probabilities in a Single Forward Pass

Traditional generative models follow the "guess the next word" route, with each sampling carrying inherent randomness. d1-ultra's design goal, however, isn't to write articles or chat — it's to produce a probability judgment that can be used directly for decision-making. "Single forward pass" means the model doesn't need multiple samplings, voting, or chain-of-thought reasoning; it directly emits a probability result, dramatically reducing inference latency and computational overhead.

"Calibrated probability" is the key here. A well-calibrated probability model, when it says "70% chance of rain," roughly corresponds to a 70% real-world chance of rain. For scenarios like risk control, medical triage, and industrial quality inspection, outputting trustworthy probabilities is far more valuable than outputting fluent text. d1-ultra makes this its core selling point, and the direction is clear.

Calibration is a core concept in the field of probabilistic forecasting, measuring the degree of consistency between a model's output confidence and its actual accuracy. Commonly used quantitative metrics include Expected Calibration Error (ECE) and Brier Score. A perfectly calibrated model should have an actual positive rate approaching p among all samples it predicts with "confidence p."

Modern deep neural networks generally suffer from overconfidence — Softmax output probabilities tend to be inflated, so when a model claims 95% confidence, its actual accuracy might only be 80%. Common calibration methods include Temperature Scaling, Platt Scaling, and Isotonic Regression, which typically fit a post-processing mapping on a validation set to correct raw probabilities.

d1-ultra claims to output calibrated probabilities in a single forward pass, meaning it may have embedded the calibration mechanism directly into the training objective or output layer design, rather than relying on post-hoc correction. This is significant for latency-sensitive scenarios (such as real-time risk control), because post-processing calibration itself introduces additional computational steps.

Single Forward Pass stands in contrast to multiple sampling (e.g., Self-Consistency) and chain-of-thought reasoning. When current mainstream large language models need to make judgments, common approaches include having the model generate multiple reasoning paths and then vote for the majority, or repeatedly sampling and tallying frequencies to approximate probabilities. While effective, these methods' inference costs grow linearly with the number of samplings.

d1-ultra's design approach is to complete the "input → probability" mapping in a single computation, eliminating the sampling and aggregation steps. This not only reduces per-inference latency and compute overhead but also makes the output deterministic — the same input always yields the same probability value, which is especially important for decision scenarios that require auditing and reproducibility (such as compliance risk control).

Differentiation Through Multimodal Input

Dual text and image input allows d1-ultra to handle tasks such as "the probability that this X-ray shows an abnormality" or "the likelihood that this scanned contract contains risky clauses." Compared to text-only decision models, multimodal decision-making covers a broader range of real-world business cases, since many real decisions are based on images, scans, or screenshots.

It's worth noting, however, that the source material only provides the model's basic positioning and score — it doesn't disclose architectural details, training data composition, or evaluation methodology. Liquid AI previously gained recognition for its Liquid Neural Networks research, but whether d1-ultra continues this technical approach cannot be confirmed from publicly available information.

AA Decision Index Score of 65 and Pricing Positioning

The Artificial Analysis Decision Index is a decision-capability evaluation metric from third-party evaluation firm Artificial Analysis. d1-ultra scored 65 — where this falls in the current decision model landscape requires looking at the full leaderboard; a single score isn't enough to judge.

On pricing, at $0.042 per million input tokens, this is on the low end among decision models. The combination of low input cost and single forward pass gives d1-ultra a cost advantage in high-concurrency, high-volume decision scenarios — such as e-commerce moderation systems that need to classify massive volumes of product images.

Artificial Analysis is an independent AI model evaluation organization known for conducting standardized benchmark tests on language models and publishing leaderboards. Its evaluation dimensions typically cover reasoning ability, knowledge level, cost efficiency, and more. Decision Index is its specialized evaluation metric for decision-oriented tasks, but the index's specific evaluation set composition, task type distribution, and scoring methodology are only limitedly disclosed in public information.

When interpreting such third-party scores, several background factors need attention: difficulty distributions vary greatly across evaluation sets, so 65 might mean leading on one benchmark but merely mid-tier on another; moreover, decision-oriented evaluations often involve choices under uncertainty, where the very definition of a "correct answer" is more ambiguous than in generation tasks. Therefore, when comparing 65 against competitors, priority should be given to confirming whether the evaluation set covers decision types similar to one's own business.

The Reality and Challenges of the Decision Model Track

Carving "decision-making" out of generative models into a separate category has been a clear trend in the AI application layer over the past year or two. Generative models excel at expression, but their output probability distributions aren't inherently calibrated — using them directly as decision bases carries systematic bias risk. Purpose-trained decision models attempt to fill this gap.

The challenges facing d1-ultra are equally obvious: evaluating decision tasks is inherently harder to standardize than evaluating generation tasks. Generation quality can be measured through human scoring or benchmark tests, but whether a "decision" is correct often depends on downstream real-world outcomes, with long evaluation cycles and high noise. Behind the number 65 — which decision types the evaluation set covers, and whether it includes real business distributions — all affect its reference value.

For developers, d1-ultra is worth attention because it offers a decision-making path that doesn't rely on large language model generation capabilities, with controllable costs and native image input support. If your business has substantial "judgment" rather than "generation" needs, this type of model may be more suitable than general-purpose large models.

Liquid Neural Networks is Liquid AI's founding technology track, originating from MIT CSAIL research. Its core idea is to use dynamically changing differential equations to describe neuron states, enabling the network to adjust its own dynamics in real time based on input during inference. Compared to traditional Transformers' fixed weights, liquid networks have theoretical advantages in temporal modeling and adaptive control, and can be made with very small parameter counts.

Whether d1-ultra continues this architecture cannot currently be confirmed, but this background is worth noting: if d1-ultra is indeed based on liquid networks or their variants, its performance on temporal decision tasks (such as sensor streams, real-time risk control) may outperform Transformer models of equivalent scale; if it actually uses mainstream Transformer architecture, there may be narrative tension between Liquid AI's brand positioning and its technical track. Clarifying this point will directly affect developers' judgment of its applicable scenarios.

Aspects to Watch Going Forward

Currently, information about d1-ultra remains concentrated on three points: positioning, score, and price. The model's actual calibration performance, the specific effects of multimodal fusion, and its generalization ability on real business distributions all still need more third-party testing to verify. Whether Liquid AI will publish a technical report or model card will determine whether this product can be more broadly evaluated and adopted.

The decision model direction has considerable imaginative potential, but it's still some distance from becoming a mature category. d1-ultra is another attempt on this path, and its actual performance is more worth tracking than a single score at launch.

One-line Summary

Liquid AI releases multimodal decision model d1-ultra, scoring 65 on the AA Decision Index, outputting calibrated probabilities in a single forward pass, with input costs of just $0.042 per million tokens.

Paragraph Summary

Liquid AI has launched d1-ultra, positioned as a decision model with dual text and image input, differentiating itself from mainstream conversational generation models. Its core feature is directly outputting calibrated probabilities in a single forward pass, eliminating multiple sampling and chain-of-thought reasoning, significantly reducing latency and compute overhead — suitable for scenarios like risk control, medical triage, and industrial quality inspection that require trustworthy probabilistic judgments. The model scores 65 on the Artificial Analysis Decision Index, with input costs of $0.042 per million tokens, giving it a cost advantage in high-concurrency batch decision-making. However, the source material doesn't disclose architectural details, training data, or evaluation methodology, and actual calibration performance and generalization ability still require third-party verification. Standardizing decision model evaluation is highly difficult, and a single score isn't enough to measure real business value.