Liquid AI Releases d1 Decision Model With Zero Output Tokens

Liquid AI has released d1, a novel decision model designed for structured decision-making rather than text generation. This model accepts context and a set of typed questions, returning calibrated probabilities across a predefined set of outcomes in a single API call, crucially with zero generated output tokens. The intended applications for d1 are tasks that are currently often handled by general-purpose large language models (LLMs), including classification, ticket routing, scoring, moderation, and LLM-as-judge evaluations. The d1 model is available today as a hosted API through the Liquid API under the model name d1:free. Liquid's model library specifies that d1 is API-only and not trainable, meaning there are no downloadable weights in formats like GGUF, MLX, or ONNX for self-hosting. A decision model, as defined by Liquid AI, evaluates a given situation and provides a typed answer from a set of options that are specified prior to the API call. Unlike LLMs, it does not generate new text. A key metric highlighted by Liquid AI is that usage.output_tokens is consistently zero for every response. The company provides a simple rule for migration: if the required answer is one of a known set of N options, a decision model like d1 is suitable; if the model needs to compose a new string, then an LLM should be retained. The d1 model operates using three fundamental primitives: Noul, Choice, and Score. Noul is used for binary yes/no questions, returning a probability between 0 and 1. For instance, Liquid AI provides an example where the question 'Is this message a complaint?' returned a probability of 0.999. Choice is designed to select one option from a named set, returning the top-ranked pick, the complete probability distribution across all options, and a confidence value. An example cited is a double-charge ticket scoring 0.9997 on the 'billing' category. Score is utilized to rate input against an ordered rubric, returning a probability-weighted position. Levels in the rubric are indexed starting from 0; for example, a 4-level urgency rubric would span from 0 to 3. A production outage, in this context, scored 2.9995. These three question types can be mixed within a single API request, with the model evaluating all questions against the same underlying state in a single round trip. An API call to d1 involves three components: the model identifier (d1), the state (which can be plain text or a JSON object representing the context), and the questions to be answered. These requests are directed to the endpoint POST https://api.liquid.ai/decisions/v1/systemone. API keys are managed through the console.liquid.ai portal and begin with the prefix 'liquid_'. Liquid AI offers client SDKs for integration, including TypeSafe AI’s typesafe-sdk for Python and @typesafe-ai/sdk for TypeScript.