Liquid AI released d1-3B and d1-omni-600M open-weight decision models. d1-3B scores 48.57 on Decision Index v0.2.1. It runs on NVIDIA DGX to Jetson hardware. d1-3B answers a question in 8 ms on RTX 4090. It answers in 16 ms on Jetson AGX Thor. It answers in 26 ms on Jetson AGX Orin. It answers in 50 ms on Jetson Orin Nano. d1-omni-600M handles text, image, and audio inputs. It scores 15.95 on the same index. Both models are available on Hugging Face. They support llama.cpp across Apple, AMD, Qualcomm, and NVIDIA. Demos run d1-3B in real time on live camera input.
What changed
Decision models required token generation and multi-step reasoning.
What it unlocks
Making single-pass decisions on text, vision, and audio at edge latency.
- 48.57 Decision Index v0.2.1 score for d1-3B
- 15.95 Decision Index v0.2.1 score for d1-omni-600M
- 8 ms latency on RTX 4090 for d1-3B
- 16 ms latency on Jetson AGX Thor for d1-3B
What you need to act on it
- NVIDIA or Apple hardware
- llama.cpp support
Sources