Liquid AI Releases Open-Weight d1 Decision Models with Ultra-Low Latency Inference
Liquid AI has released two open-weight decision models, d1-3B and d1-omni-600M, offering multimodal capabilities across text, image, and voice inputs. The models achieve decision latencies as low as 8 milliseconds on desktop GPUs like the NVIDIA RTX 4090. By publishing lightweight, high-speed decision models openly on Hugging Face, Liquid AI enables developers to deploy real-time intelligence directly onto edge devices and robotics. This reduces reliance on costly cloud infrastructure for applications requiring instant multimodal reasoning. The 3.12-billion-parameter d1-3B model scored 48.57 on the Decision Index v0.2.1 benchmark, outperforming existing models under 10 billion parameters. On edge computing hardware, d1-3B recorded single-decision latencies of 16 milliseconds on NVIDIA Jetson AGX Thor and 50 milliseconds on Jetson Orin Nano.
## BACKGROUND
Liquid AI is a company focused on dynamic neural architectures capable of efficient computing on continuous-time data streams. Decision engines evaluate dynamic inputs and clear criteria to select actions, making them critical for high-speed edge tasks like drone navigation, robotics control, and interactive interfaces.