- Published on
Liquid AI Releases d1: A Decision Model That Returns Calibrated Probabilities With Zero Output Tokens
- Authors
- Name
- DeepHorus
Liquid AI has released d1, a decision model designed to turn context and predefined questions into typed probabilities rather than generated prose. Available now through the Liquid API as “d1:free,” it targets classification, ticket routing, scoring, moderation, reranking, and other “LLM-as-judge” workloads. Its defining technical claim is simple: every response reports zero output tokens because the system selects among fixed outcomes instead of decoding a new string. That makes d1 an alternative for high-volume decision APIs, not a replacement for general-purpose language models.
Most teams currently ask a large language model to return JSON containing a label, category, or score. Even when the answer is one word, the model must generate output tokens, and developers must handle malformed schemas, retries, inconsistent confidence, and several network round trips when evaluating multiple questions. d1 is built around the narrower rule that if the result is one of N known options, a decision model is more appropriate; if the system must draft, summarize, or compose original text, a conventional LLM remains the right tool. This division aims to move routine judgment work away from expensive decoding paths without asking teams to redesign their broader AI stack.
d1 exposes three primitives. Noul asks a yes/no question and returns a probability from zero to one; Liquid’s sample assigns 0.999 to a complaint message. Choice selects from a named set and supplies the winner, the full distribution, and confidence, as shown by a 0.9997 “billing” score for a double-charge ticket. Score places an input on an ordered, zero-indexed rubric, with a four-level urgency scale running from zero to three; in one example, a production outage receives 2.9995. Developers can mix all three in a single request, allowing the model to assess several questions against the same state in one round trip.
API requests contain the model name, a state represented as plain text or JSON, and a typed question object. Calls are sent to the SystemOne decisions endpoint, authenticated with a “liquid_” key, and supported by TypeSafe AI’s Python and TypeScript SDKs. Because each primitive has a fixed return shape, Liquid says applications can avoid malformed JSON and retry logic while gaining probability estimates that are easier to threshold than a model’s self-reported confidence. The hosted-only model is listed as non-trainable, however, and Liquid is not publishing GGUF, MLX, or ONNX weights.
The main operational appeal is cost and predictability: users are not billed for generated output tokens, and latency does not expand with the length of a decoded answer. Consolidating three sequential classifications into one call can also reduce orchestration overhead. Liquid’s moderation example blocks inputs above 0.8, permits those below 0.2, and routes the uncertain middle to human review, while a support example escalates to a more capable model when router confidence falls below 0.5. These patterns could make d1 useful for real-time systems where a stable contract and fast fallback policy matter more than conversational fluency.
Reaction to the launch is likely to center on the contrast between a narrow API and the flexibility of general LLMs, especially because the free endpoint is immediately accessible while self-hosting is not. Liquid claims calibrated probabilities and more consistent repeated evaluations, but those claims still require independent testing on each customer’s data, labels, and threshold-sensitive edge cases. Probability outputs are not proof that a model is correct, and teams will still need monitoring, drift detection, privacy review, and human escalation for high-stakes decisions. The absence of downloadable weights may also limit adoption among organizations with strict data-residency or air-gapped infrastructure requirements.
Liquid’s Road Decider demo illustrates both the promise and the practical lessons of the model. The browser-based racer asks d1 to choose left, center, or right two to five times per second, with a Node.js and Vite proxy keeping the API key off the client; a comparison mode pits it against TypeSafe’s jev-1.13 through OpenRouter. Notably, compact per-lane state summaries produced stronger decisions than feeding the model a raw road grid, suggesting that decision-model performance still depends heavily on representation engineering. Going forward, Liquid could expand d1 through more primitives, higher-throughput tiers, private deployment options, and stronger calibration evidence, but its future depends on proving that specialized decisions can outperform prompt-based LLM workflows at comparable accuracy, latency, and total cost.
