Liquid AI Releases d1: A Decision Model With Zero Output Tokens
Liquid AI launches d1, a decision model via hosted API that returns calibrated probabilities with zero output tokens for classification tasks.
Stock photo for illustration only, not from the actual event
- Liquid AI has released d1, a new decision model available as a hosted API.
- It returns typed answers from predefined options with zero output tokens.
- Supports evaluating multiple questions against the same state in a single round trip.
- Designed to replace LLMs for classification, moderation, and routing tasks.
Liquid AI has officially introduced d1, a specialized decision model engineered to evaluate situations and return typed answers from predefined options. A core characteristic of this model is its non-generative nature, ensuring that usage.output_tokens remains at zero for every single response.
The d1 model is accessible today as a hosted API under the model name d1:free via the Liquid API. According to Liquid's model library, it is strictly API-only and non-trainable, meaning there are no GGUF, MLX, or ONNX weights available for self-hosting.
Stock photo for illustration only, not from the actual event
Each API call to d1 consists of three primary components: the model identifier, the state (provided as plain text or a JSON object), and the questions. Requests are directed to https://api.liquid.ai/decisions/v1/systemone using authentication keys generated from console.liquid.ai starting with liquid_. Supported clients include TypeSafe AI's typesafe-sdk for Python and @typesafe-ai/sdk for TypeScript.
The introduction of non-generative decision models represents an efficient paradigm shift in AI infrastructure. By bypassing natural language generation entirely when only categorical answers are needed, systems can achieve higher consistency, lower latency, and reduced computational overhead compared to traditional Large Language Models.
Liquid AI's migration guide outlines a straightforward rule for developers: if the answer corresponds to one of N known options, a decision model should be utilized. If the system requires composing entirely new strings, retaining a standard LLM is advised. Furthermore, users can mix all three types within a single request, evaluating multiple questions against the same state in one round trip.
"if the answer is one of N known options, use a decision model. If the model must compose a new string, keep your LLM."
Liquid AI Migration Guide
Migrating classification workloads from LLMs to d1 offers several distinct advantages:
- Practical Thresholds via Probabilities: Moderation examples can block inputs above 0.8, allow inputs below 0.2, and route intermediate bands to human reviewers.
- Reliable Model Routing: Fallbacks to more capable model tiers trigger automatically when router confidence drops below 0.5.
- Consistent Evaluations: Repeated evaluations of identical inputs yield higher consistency, minimizing verdict flips.
Developers should continue using traditional LLMs for summarization, content drafting, multi-turn chat, code generation, and complex multi-step reasoning. All data was checked against each vendor's primary page on September 29, 2026.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment