Liquid AI has launched d1, a decision model designed to evaluate situations and provide typed answers based on predefined options. Unlike traditional large language models, d1 does not generate text, resulting in zero output tokens per response. The model is currently available as a hosted API under the name d1:free.
The model operates via the Liquid API, with requests consisting of the model name, state, and specific questions. Integration is supported through the TypeSafe AI SDK for Python and TypeScript. Developers can mix multiple question types in a single request, with the model evaluating all inputs against the same state in one round trip.
Liquid AI suggests using d1 for classification tasks where the answer is one of several known options, while reserving LLMs for summarization, code generation, and complex reasoning. The model provides calibrated probabilities, allowing users to set thresholds for automated actions, such as moderation or routing, and offers higher consistency in repeated evaluations.
To demonstrate the model's performance, Liquid AI developed a survival racer game where d1 makes steering decisions 2 to 5 times per second. The implementation uses Node.js 18+ and a Vite proxy to secure API keys. The company notes that d1 is not self-hostable, as there are no GGUF, MLX, or ONNX weights available for local deployment.
