AI news · October 10, 2026
Liquid AI ships open d1 decision models for edge text, vision, and audio
Liquid AI's open d1-3B answers a decision in 8ms on an RTX 4090 and under 50ms on measured Jetson devices; d1-omni-600M adds text, image, and audio inputs.
- d1-3B parameters
- 3B
- RTX 4090 decision time
- 8ms
- d1-omni parameters
- 600M
TLDR AI's Oct 8 issue featured Liquid AI's d1 release. Liquid's Oct 7 post describes d1-3B as an open-weight model that returns a structured answer in one forward pass rather than generating tokens; it scores 48.57 on the public Decision Index and runs in 8ms on an RTX 4090, 16ms on Jetson Thor, 26ms on Jetson Orin, and 50ms on Orin Nano. The d1-omni-600M checkpoint adds text, image, and audio inputs.
Both models are available on Hugging Face with llama.cpp support. That makes d1 a different choice from Microsoft's hosted Decision-1: d1 targets local and edge use, while Decision-1 offers a managed API and calibrated choice scores. Liquid reports its own benchmarks, and the omni model is experimental, so a small local test matters more than the headline speed.
What you can do with it
Download d1-3B or d1-omni-600M from Hugging Face and start with routing, moderation, or camera decisions. Measure accuracy and time on your device before replacing a hosted classifier; the models are open-weight, but the omni checkpoint is experimental.
Our take
This is a practical release because it targets the small decisions that add delay to an agent loop, and the reported 8ms and 50ms results fit interactive systems. The catch is quality: Liquid reports its own tests, and fast scores can still fail on your labels.
Links Release detailsDownload d1-3B
Source: Liquid AI — Made With Models writes the brief; the reporting is theirs.