Liquid AI’s d1 decision model solves a common pain point for teams that still rely on general language models for structured tasks like classification, ticket routing, scoring, moderation, reranking and LLM‑as‑judge checks. Instead of prompting an LLM to generate a label and then parsing its output, d1 returns calibrated probabilities for a predefined set of outcomes in a single API call with zero output tokens. This eliminates the cost and variability associated with token billing, removes the need for JSON schema validation, and provides predictable latency because there is no decoding loop that grows with response length.
The model works with three primitive question types. A Noul question yields a yes/no probability, useful for binary moderation thresholds. A Choice question selects one option from a named set and supplies the full distribution plus a confidence score, ideal for routing tickets to the right team or model tier. A Score question rates input on an ordered rubric, delivering a continuous probability‑weighted position that can trigger escalation workflows. All three types can be combined in one request, so a single state description can drive multiple decisions at once.
Deploying d1 is straightforward: it is available as a hosted API under the model name d1:free. Authentication uses a Liquid API key, and official SDKs are provided for Python and TypeScript. Because the model is not trainable, there are no weights to manage, no GGUF, MLX or ONNX files to host, and no need to worry about version drift. Teams can replace repetitive LLM classification calls with a single d1 invocation, cut costs, reduce latency spikes, and gain reliable uncertainty estimates that make threshold‑based actions—like blocking content, escalating tickets, or falling back to human review—simple and consistent.
In practice, a ticket such as “I was charged twice for my subscription last month” can be sent to d1 with a Choice question about the primary topic. The model instantly returns the probabilities for billing, account, returns, shipping and technical, together with a confidence value, allowing an automated router to send the case to the billing specialist without any post‑processing. Similar flows apply to moderation, urgency scoring, and reranking, delivering a cleaner, cheaper and more predictable alternative to generic LLMs for structured decision‑making.
#AI #Product #DecisionModels #LLM #Automation #MLOps