Liquid AI Releases Decision Model d1

Liquid AI has released d1, its first Decision Model specialized for classification, routing, and scoring rather than text generation. This model uses the same interface as TypeSafe AI's System One model (Jev), with output tokens fixed at always 0, meaning cost and latency are determined by input length.

Key Features and Usage Guide

• Differentiation from LLMs: Use d1 when the answer is one of a fixed set of choices, and use an LLM when new string generation is required. d1 offers advantages such as low latency, Calibrated Probability, no schema errors, and parallel evaluation.
• Question Types: Supports three types: Noul, which returns yes/no probabilities; Choice, which returns category selection and confidence; and Score, which returns ordinal scale scores.
• Confidence Calculation: The confidence in Choice responses is calibrated based on the number of options (K) and the top-1 probability (p_max), rather than just the top-1 probability. A larger number of options results in a lower confidence value even with the same top-1 probability.

Performance and Benchmark Verification

Liquid AI claimed that d1 is the first model to surpass Jev on the Hugging Face Decision Index, but independent reproduction results showed a total score difference of only 1.0 point (d1 58.9 vs Jev 57.9). While it showed superiority in the Arts domain, it showed inferiority in the Knowledge domain. On external benchmarks (typed-decisions), d1 demonstrated better accuracy (0.742 vs 0.727), KL divergence, and ECE metrics compared to Jev.

Considerations for Adoption

Currently, d1's weights are not public and it is only available via the Liquid API. Fine-tuning or on-device execution is not supported, and specific operational information such as pricing, usage limits, and supported language lists is not documented. Korean language reliability has also not been verified.