Evidence at a glance
The mechanism in one line
Compress the visual or contextual input before the main reasoning path.
Route or verify the expensive step instead of repeating the full path.
Translate the mechanism into a bounded deployment or evaluation check.
From Generating Answers to Returning Decisions
Liquid AI released d1-ultra on October 15, 2026, positioning it as a text-and-image model for more complex decisions. It is not aimed at chat, summarization, or code generation. Instead, it targets bounded tasks such as classification, routing, and scoring: a developer supplies context and a question, and the model returns answer probabilities rather than a block of prose.
The release matters to technical leaders because it changes where the model sits in an application pipeline. A conventional generative interface produces text that rules, a parser, or another model must turn into an action. d1-ultra aims to hand a decision directly to software. Removing the text-generation step can simplify the interface, but probability-shaped output does not by itself make a decision reliable.
The Question Type Defines the Interface
A d1-ultra request contains a state, optional images, and one or more typed questions. The model evaluates those questions in a single forward pass. The documentation describes three forms: Noul for yes-or-no judgments, Choice for probabilities across a set of labels, and Score for placing a result on an ordered scale. Several questions can use the same context, avoiding a separate text-generation call for every judgment.
A quality-control system, for example, could ask whether an item is defective and then choose a defect category. A claims workflow could provide a photo and description and ask whether the damage meets a particular rule. Fixed answer structures make it easier to connect results to thresholds, routing, and review queues. The trade-off is that teams must first translate real business questions into supported decision types. The interface is not a substitute for open-ended explanations or knowledge of the surrounding workflow.
What the Score of 65 Does—and Does Not—Show
Liquid AI reports that d1-ultra scored 65 and ranked first on the Artificial Analysis Decision Index. This is not a single-task accuracy score. The index gives equal weight to four evaluation areas: knowledge and hallucination, reasoning, tool calling, and safety classification, each worth 25 percent. Knowledge and hallucination are split equally within their category. A score of 65 therefore means a composite index result, not that the model answers 65 percent of questions correctly.
The source and timing of the evidence also matter. Liquid AI says d1-ultra beat the earlier d1 and Jev 1.14 across its 112 evaluations, but the announcement does not publish the full task-by-task results, so that comparison cannot be independently checked from the available material. Artificial Analysis’s October 13 article reported Jev 1.13 at 58 as the leader at that time. That earlier table should not be treated as a contemporaneous post-launch comparison with d1-ultra. The index offers an external evaluation framework, while the vendor’s more detailed comparative claims still need fuller disclosure.
No Output Tokens Does Not Mean No Cost
d1-ultra is available through the Liquid API at a published price of $0.042 per million input tokens, and responses contain zero output tokens. Images are counted in 32-by-32-pixel patches, at 1.5 tokens per patch. Each question counts the attached images again. If one image supports several separate judgments, its input cost is therefore counted repeatedly. “No output tokens” alone is not enough to estimate the total bill.
The deployment model is another important trade-off. Liquid says d1-ultra uses a larger open-weight backbone than d1 with Liquid-specific modifications, but has not disclosed the backbone, parameter count, or context length. The d1-ultra model itself is currently offered only through a hosted API, with no downloadable weights or quantized variants announced. The d1-3B and d1-omni-600M models released earlier are open-weight products and should not be conflated with d1-ultra. Teams can integrate the hosted service, but the available material does not establish whether local deployment, model migration, or independent reproduction is feasible.
Validate the Probabilities Before Acting on Them
Liquid AI describes the probability outputs as specially trained to be more reliable, but the public material does not provide calibration error, test-set details, or a complete comparison. That gap matters in business systems: a probability supports a useful threshold only when it corresponds to observed frequencies and remains stable on the target data. A value of 0.85 should not automatically be read as an 85 percent chance that the judgment is reliable in a particular workflow.
A safer integration path is to start with a bounded task whose outcomes can be checked, use your own data to examine error patterns and probability calibration, and then decide which ranges can be automated and which should go to human review. Routing and moderation may offer lower-risk starting points, but workflows with financial, quality, or safety consequences need review thresholds and fallback paths. d1-ultra’s value is a more direct interface for structured decisions. Whether it reduces total cost or risk depends on the combined effects of task-specific evaluation, repeated image billing, and reliance on a hosted service.