Liquid AI has introduced d1, a specialized decision model for production software, according to AlphaSignal. Instead of generating text for an application to parse, it returns typed answers with probabilities and reports zero generated output tokens—a design that can reduce response delays and parsing failures in supported classification, routing, approval and scoring tasks.
A decision interface rather than a text response
AlphaSignal describes a different output path from the familiar process of asking a language model to write an answer and then converting that text into a required schema. d1 supplies the typed result directly. Its potential benefit is specific to workloads whose answers fit the supported response types, rather than tasks that require an unrestricted written response.
The interface also exposes numerical probabilities, AlphaSignal reports, rather than requiring a model to express its confidence in words. Liquid AI describes those probabilities as calibrated. That company claim concerns the probability values returned by the API, not a confidence label generated as part of a chat response.
Noul, Choice and Score in one request
According to AlphaSignal, each request supplies a state in plain text or JSON, together with one or more questions. Each question specifies an answer type. The same state can support a mixture of types within one call: a customer-support application could request the message’s intent, its urgency and the probability that it reports a bug, instead of making three separate model calls.
For binary questions, AlphaSignal describes Noul as returning a probability between 0 and 1. Its stated uses include approval steps and flags, with a threshold determining a yes-or-no result. Choice handles named alternatives instead: it returns a probability distribution across those options, accompanied by a confidence measure, for classification, routing or action selection.
Score addresses ordered evaluations, according to AlphaSignal, by returning a probability-weighted position on a rating scale. The value can sit between levels: a four-level rubric might produce 2.9995 rather than an integer. An application can retain that value, assign it to a category or compare it against a threshold.
Where structured decisions fit
AlphaSignal identifies message and ticket routing as potential uses for d1, alongside content moderation and approval gates for agent tool calls. Other described applications include choosing between models, evaluating risk or priority, and reranking RAG results using relevance probabilities.
Generative work is outside d1’s scope. As reported by AlphaSignal, Liquid AI recommends a language model for free-form writing and summarization, as well as code generation and multi-turn conversations. Open-ended questions and complex, multi-step reasoning also belong with a language model; producing a new sentence requires either a generative model or a separate templating layer.
A more visual illustration comes from Liquid AI’s Road Decider demonstration, described by AlphaSignal. In the pixel-art driving game, d1 acts as the controller and chooses LEFT, CENTER or RIGHT for each frame, returning a confidence score with the selection. The example demonstrates the decision interface through a game rather than a deployed driving system.
Reported changes from earlier decision models
Liquid AI claims four improvements over its earlier decision models, AlphaSignal reports. Two concern the inputs it handles: higher scores in multilingual evaluations and better processing of longer inputs. The company also reports faster structured decisions in software workflows and greater resistance to prompt injection specifically within the state field supplied to the model.
The Jev Decision Index comparison
Liquid AI claims d1 has taken first place in the Jev Decision Index, according to AlphaSignal. The evaluation uses a fixed collection of 132,422 requests across 37 benchmarks. AlphaSignal identifies Jev 1.13 as the previous leader, with an overall score of approximately 74.4 before d1 reached the top position.
AlphaSignal explains that 19 of the 37 benchmarks contribute to the headline composite. Five areas carry equal weight: Tools & Automation and Retrieval & Classification are joined by Language Understanding, Knowledge & Reasoning, and Arts & Human Judgment. The final index ranges from 0 to 100 and is calculated as 100 times the mean of the five area scores.
Access through Liquid API
d1 is available through Liquid API under the model name d1:free, with a free tier, according to AlphaSignal. Developers need a Liquid API key to access it.
