Liquid AI Releases d1-3B Decision Model for Edge Devices

On October 7, 2026, Liquid AI released the weights for d1-3B and d1-omni-600M, decision models that return probabilities for yes/no, multiple-choice, and scoring questions in a single forward pass without generating text on edge devices. These models support Transformers and llama.cpp, enabling local execution.

Model Architecture and Performance

d1-3B is a decoder model based on LFM2.5-VL-3B, supporting text, JSON, and image inputs. It achieved a score of 48.57 on the Decision Index 0.2.1, recording the highest performance among models under 10B, and surpassed Decider 4B (81.1 points) with an average score of 82.9 across seven public benchmarks. The latency for processing a single question on Jetson AGX Thor is 16ms.

d1-omni-600M is an early-stage research model based on a bidirectional encoder, supporting text and image or audio inputs. With a Decision Index score of 15.95, it is better suited for narrow-domain routing than general-purpose use and is currently in the experimental phase.

License and Caveats

The license is LFM Open License v1.0, which restricts commercial use by companies with annual revenue exceeding $10 million. The existing hosted API d1 (Decision Index ~60) and the newly released models are separate, with the released models showing performance approximately 9 points lower than the API. Additionally, Korean speech support has not been verified, so confirmation is required before adoption.