Liquid AI Releases Open d1 Edge Decision Models

Summary

Liquid AI has released d1-3B and d1-omni-600M, two open-weight decision models designed to produce an answer in a single forward pass rather than generate tokens. The 3B model accepts text and images, while the experimental 600M model accepts either text and images or text and audio. On Decision Index v0.2.1, d1-3B scores 48.57, leading models below 10B parameters and matching the reported score of the much larger Decider 35B-A3B; d1-omni-600M scores 15.95. Across seven public text benchmarks, d1-3B records a mean of 82.9, ahead of Decider 4B at 81.1, while d1-omni-600M reaches 78.4 and exceeds Decider 2B at 77.1. The models use different backbones and staged multimodal training, including weight averaging, adapters, and LoRA updates. Liquid AI reports d1-3B latency as low as 8 ms for one question on an RTX 4090 and 16 ms on Jetson AGX Thor; it remains below 50 ms on each measured edge device for a single question. The company does not report the private vision split of Decision Index v0.3 and says dedicated audio decision benchmarks remain an open problem. Both models are available on Hugging Face, with llama.cpp support and demos for live camera inputs and Isaac Sim navigation.