Liquid AI Launches d1 to Make Decisions Without Tokens

Liquid AI has released its first dedicated decision model, d1, designed to handle classification, routing, and scoring tasks within production software. The model recently secured the top position on Hugging Face's Jev Decision Index, overtaking the previous leader, Jev 1.13, which held a composite score near 74.4. The Jev index evaluates performance using a frozen dataset of 132,422 requests spread across 37 benchmarks, with 19 specific benchmarks determining the final composite score.

Unlike standard language models that generate text sequentially, d1 delivers structured answers alongside calibrated probabilities without producing any output tokens. By bypassing the autoregressive decoding phase, the system avoids typical integration hurdles such as corrupted JSON structures, absent fields, or invalid enum parameters. This architecture helps developers reduce latency, minimize parsing failures, and eliminate output-token charges for structured workflows.

Developers can interact with the model using three distinct primitives: Noul for binary yes-or-no queries, Choice for picking from a list of options, and Score for grading on a continuous scale. A single API request to the endpoint at https://api.liquid.ai/decisions/v1/systemone can combine all three primitives against a single input state. For example, a customer service system can simultaneously evaluate a message's intent, urgency score, and bug-report probability in a single round trip.

The model is currently accessible through the Liquid API console under the name d1:free, with integration for OpenRouter planned for a later date. To demonstrate its real-time capabilities, Liquid AI showcased a Road Decider driving game, where the model serves as a controller to choose between steering left, right, or remaining in the center for each frame. While d1 offers improved resistance to prompt injection, Liquid AI advises developers to maintain external validation and keep generative workloads on traditional language models.