Liquid AI unveils open-weight multimodal decision models for real-time applications

Liquid AI has launched two open-weight multimodal models, d1-3B and d1-omni-600M, designed for real-time decision-making. These models, part of the d1 decision model family, process text and images or text with audio, but do not generate text outputs. They provide calibrated answers in a single pass with zero output tokens, suitable for deployment on NVIDIA hardware including DGX servers, RTX workstations, and Jetson edge boards. The d1-3B model, with 3.12 billion parameters, outperforms other models under 10 billion parameters and scores highly on text benchmarks. d1-omni-600M, with 587 million parameters, excels in specific tasks like Civil Comments and PAWS-X. Both models are available on Hugging Face and support llama.cpp, with commercial use permitted under the LFM Open License v1.0. Latency tests show d1-3B operates at 8 ms on an RTX 4090, while Jetson devices range from 16 ms to 50 ms. Liquid AI recommends float16 for d1-omni-600M on GPUs. The models are intended for routing, moderation, and intent classification, with no chat capabilities.