Liquid AI d1-ultra Lands on AI Gateway: A Multimodal Decision Model with Image Detection Support

When LLMs Stop Just "Talking": The Rise of Decision Models

Most people picture large language models as text generators — writing articles, answering questions, writing code. But a different class of models is taking a completely different path: generating no text at all, only outputting structured probability judgments. Liquid AI's d1-ultra is exactly this kind of outlier, and it has just become available through AI Gateway.

According to AI Gateway's official information, d1-ultra is a multimodal decision model purpose-built for classification, routing, and scoring tasks, with support for both text and image inputs. Rather than asking it to "write an analysis," you can simply ask it: "Is the screen in this product photo cracked?" It returns a probability value, not a paragraph of explanation.

This design philosophy means d1-ultra is not positioned to replace chatbots, but to be embedded at the decision nodes of automated workflows.

Three Ways to Ask, One Core Logic

Boolean, Choice, Score

d1-ultra's interaction model revolves around three question types:

• Boolean: Pose a yes/no question and the model returns the probability of "yes." For example, judging whether a customer's description indicates the item arrived damaged.
• Choice: Select among predefined named categories, suited for multi-class scenarios.
• Score: Rate on an ordinal scale, ideal for risk assessment or quality grading.

The key is that you can ask multiple questions about the same input within a single request, and use the returned probabilities to decide when to act automatically and when to escalate to human review. This is far more flexible than simply calling a classification API.

Why structured Q&A beats text generation for decisions

The traditional approach has a generative LLM output a passage of natural language, then uses regex or a second model to parse out "yes/no." The problem with this pipeline: unstable formatting, high parsing cost, and lost probability information. The Boolean, Choice, and Score question types essentially turn a "decision" into a single forward inference, where the output is directly a numerically consumable value.

This design approach aligns with the recent rise of LLM-as-a-judge and logit-based classification, but goes further: it doesn't borrow the output distribution of a generative model — it's optimized for discriminative tasks at the architectural level. The Choice type corresponds to closed-set classification, the Score type to ordinal regression, and Boolean to the simplest binary classification. Together, the three cover most industrial automation decision needs.

Code invocation

Developers can call d1-ultra through three interfaces: AI SDK, an OpenAI-compatible Decisions API, and a TypeSafe-compatible API. The model name is uniformly liquid/d1-ultra.

Here is an AI SDK example that determines whether a customer reported item damage:

import { experimental_decide as decide } from 'ai';

const result = await decide({
  model: 'liquid/d1-ultra',
  state: 'The customer reports that the delivered laptop has a cracked screen.',
  questions: {
    damaged: {
      type: 'boolean',
      instructions: 'Does the customer report that the item arrived damaged?',
    },
  },
});

console.log(result.answers.damaged);

As this snippet shows, d1-ultra's usage paradigm is completely different from conventional text generation models — no flashy prompt engineering tricks, just a clear input state and structured question definitions.

The significance of AI SDK and compatible APIs

AI SDK is Vercel's open-source toolkit for calling large models, and experimental_decide sits in its experimental namespace, signaling that decision calls are a first-class citizen alongside streaming generation. OpenAI-compatible API support means existing codebases built on the openai SDK can migrate with minimal changes — just swap the base URL and model name, no need to rewrite business logic. The TypeSafe-compatible API emphasizes end-to-end type inference, so with TypeScript you can validate field names in question definitions and the shape of return structures at compile time, avoiding lost probabilities from typos at runtime.

The coexistence of three interfaces lowers the barrier to entry, letting developers choose flexibly based on their existing tech stack. This reflects the decision model's niche: not a new species that bypasses existing toolchains, but a specialized capability module embedded within them.

Calibrated Probabilities Are the Real Selling Point

The core capability Liquid AI emphasizes for d1-ultra is "calibrated probabilities." This term is often abused in AI, but here it has real meaning: the model's output probabilities should match actual frequencies. If the model says an image has an 80% probability of a defect, then across a large number of similar samples, roughly 80% should indeed have defects.

This characteristic directly determines whether automated workflows can set reliable thresholds. In a visual defect detection scenario, for example, you could set: probability above 0.95 auto-approve, below 0.6 auto-reject, and the middle band escalates to human review. If probabilities aren't calibrated, this logic breaks down.

Typical scenarios d1-ultra supports include visual defect inspection and comparing customer photos against damage claims. What these two scenarios share: the need to make fast, consistent judgments where errors are costly.

What are calibrated probabilities

Calibrated probabilities are a foundational concept in statistics and machine learning: a model's output probability values should reflect the real-world frequency of events. For example, if a model gives "70% probability of positive class" for 1,000 samples, then roughly 700 of them should actually be positive — this is called "perfect calibration." The common opposite problem is overconfidence: the model claims 90% certainty while its actual accuracy might be only 60%.

Common metrics include reliability diagrams and Expected Calibration Error (ECE). The industry often applies Platt scaling, isotonic regression, or temperature scaling to post-process and calibrate probabilities. d1-ultra's claimed calibration capability means it has undergone targeted training or post-processing, so thresholds can map directly to business rules without manually fumbling for an empirical value.

The Billing Model Is Worth Noting

d1-ultra's pricing model is quite unusual in today's API market: only input tokens are billed, no output token charges, and image inputs also count as input tokens.

Since the model generates no text — the output is just a structured probability result with negligible token consumption — not charging for output is logically consistent. For scenarios requiring heavy decision-model calls — say, processing thousands of QC images daily — this billing approach could significantly reduce costs.

As for context window, d1-ultra supports 32,768 tokens, enough to accommodate longer text inputs or multiple images.

Where Decision Models Fit in the AI Gateway Ecosystem

AI Gateway unifies d1-ultra with other models under one API. Decision calls also appear in logs and count toward the budget system. The platform bills at provider pricing with no markup and no platform fee for inference, and this rule also applies to bring-your-own-key (BYOK) requests.

From a broader perspective, the emergence of decision models answers, to some extent, the long-running debate over "whether using generative models for classification tasks is worth it." Using a GPT-class model for binary classification is both expensive and unstable, and requires parsing natural language output. The approach of specialized models like d1-ultra is to separate decision-making from generation, trading lower cost for more reliable judgments.

Of course, decision models have boundaries too. They're not suited for open-ended Q&A, content creation, or complex tasks requiring chains of reasoning. Their value lies precisely in doing one thing exceptionally well: giving trustworthy probability judgments on clearly defined questions.

Developers can try d1-ultra in AI Gateway's model marketplace, or consult the decision model documentation for more detailed question types and usage.