Liquid AI’s d1-ultra returns decisions without generating text

Liquid AI has released d1-ultra, a multimodal decision model that maps text and images to probabilities over developer-defined outcomes. d1-ultra completes each decision in a forward pass without generating any tokens.

The public Decision Index ranked d1-ultra first at publication with a score of 65, eight points ahead of the next model. For teams using chat completions for classification, moderation, routing, or scoring, the model offers a direct alternative to generating and parsing labels or JSON.

Probabilities replace generated labels

General-purpose language models produce tokens sequentially, even when an application needs only a label or score. That process adds decoding work and requires developers to validate generated output. A decision model returns structured probabilities directly, while the API response still carries the resulting numeric data.

d1-ultra accepts a state description, optional images, and one or more questions. Each question uses one of three decision primitives:

The API can batch several questions in one request, although Liquid bills each question as a separate prompt. Applications must define the outcomes and instructions in advance; the model does not produce an open-ended explanation or invent new labels.

Calibration turns confidence into a control

Calibration measures whether a model’s confidence matches its observed accuracy. Across many comparable predictions assigned a probability of 0.90, a well-calibrated model should be correct about 90% of the time. This property makes probability thresholds useful for production workflows.

Liquid says it trained d1-ultra to improve calibration as well as accuracy. A team could automatically accept decisions above a validated threshold and send lower-confidence cases to a reviewer. The threshold can also control a model cascade, allowing simple cases through while escalating ambiguous inputs to a larger model.

Production thresholds still require local validation. Class balance, image quality, language, domain shifts, and changes in user behavior can alter calibration. Teams should measure error rates at each threshold and recalibrate when the input distribution changes.

Benchmarks show a broad lead

Liquid compared d1-ultra to its previous model d1 and Jev 1.14 across 112 unique evaluations. The largest reported gains covered answer comparison, domain-specific judgment, agent tool use, and multilingual tasks. Liquid also reports higher accuracy and better calibration than Jev 1.14 at the same listed price.

The public Decision Index provides an external point of comparison, while application-specific tests remain necessary. A moderation system, visual inspection line, or medical triage workflow can have different class distributions and error costs from a general benchmark.

Input-only pricing changes the cost model

Liquid lists d1-ultra on the Liquid API at $0.042 per million input tokens, with no output-token charge. The launch post says OpenRouter and Vercel listings are planned.

Images use the same input-token rate as text. Liquid counts 1.5 tokens for each 32 × 32-pixel patch, so a 1024 × 1024 image consumes 1,536 tokens. At the listed rate, that image costs about $0.0000645 per question before text tokens are added.

Billing each question as its own prompt can materially affect multimodal workloads. Sending the same 1024 × 1024 image with ten questions may incur roughly 15,360 image tokens, plus the state and instruction tokens associated with each decision.

From image to probability

The REST API accepts a model name, shared state, optional images, and a question dictionary. This Python example submits a circuit-board defect check and handles HTTP errors:

import os
import requests

response = requests.post(
    "https://api.liquid.ai/decisions/v1/systemone",
    headers={
        "Authorization": f"Bearer {os.environ['LIQUID_API_KEY']}"
    },
    json={
        "model": "d1-ultra",
        "images": [f"data:image/jpeg;base64,{image_b64}"],
        "state": (
            "Camera image of a circuit board "
            "on the production line."
        ),
        "questions": {
            "defect": {
                "type": "noul",
                "instructions": (
                    "Does this circuit board have a defect?"
                ),
            }
        },
    },
    timeout=30,
)

response.raise_for_status()
result = response.json()
print(result)

The exact response fields and supported question options should be checked against the current decision model documentation before integrating them into a typed client or production schema.

Good fits and clear limits

The model fits workloads whose outputs can be specified before inference:

• Ticket routing, intent detection, moderation, and triage
• Quality scoring and automated answer evaluation
• Safety checks, extraction validation, and tool-call approval
• Visual inspection, image classification, and screenshot analysis
• Model routing and confidence gates for LLM cascades

Liquid’s demos include waste sorting, detecting mismatches between product photos and listing text, and selecting a game action from available choices. These examples share the same structure: unstructured multimodal input enters the model, and a predefined set of outcomes comes back with probabilities.

Free-form writing, summarization, coding, open-ended questions, and multi-turn conversation remain generative-model workloads. Applications that require rationales must obtain them separately, and generated explanations should not be treated as evidence that a classification is correct.

Test accuracy, calibration, and cost together

A production evaluation should compare d1-ultra with the existing classifier or LLM pipeline on representative data. The test should cover:

1. Task quality: accuracy, precision, recall, F1, or ranking metrics suited to the workload.
2. Calibration: Brier score, expected calibration error, and reliability plots.
3. Threshold behavior: error rates, automation coverage, and reviewer volume at each cutoff.
4. Operational performance: end-to-end latency, retries, rate limits, and failure handling.
5. Total cost: repeated image and state tokens when a request contains several questions.
6. Robustness: results by language, input quality, customer segment, and class frequency.

For structured decision pipelines currently built around chat completions, d1-ultra provides a focused comparison point: direct probability outputs, multimodal input, and input-only token billing. Its production value will depend on local calibration, threshold performance, and the cost of repeating shared context across questions.