Liquid AI open-sources Open d1: d1-3B and d1-omni-600M decision models answer in one forward pass, 8 ms on RTX 4090

Liquid AI open-sourced two models from its d1 decision family on Oct 7, 2026. Unlike generative LFMs, d1 models emit no tokens and return a calibrated answer (yes/no, choice, score) in one forward pass. d1-3B is built on the decoder-only LFM2.5-VL-3B (text+image), using weight averaging with LFM2.5-2.6B and merged seed/data-mix fine-tunes; d1-omni-600M builds on the bidirectional LFM2.5-Encoder-350M with a FastConformer audio encoder and the LFM2.5-VL-450M vision encoder. d1-3B scores 48.57 on Decision Index v0.2.1 (public split), first under 10B and on par with Decider 35B-A3B; its seven-benchmark text mean is 82.9 vs 81.1 for Decider 4B, while d1-omni-600M hits 78.4 vs 77.1 for Decider 2B. Single-question latency is 8 ms on RTX 4090, 9 ms on MI325X, 16 ms on Jetson AGX Thor and 50 ms on Jetson Orin Nano. Weights (including GGUF and w8a8) are on Hugging Face under the LFM Open License v1.0 with day-one llama.cpp and NVFP4 support. Target uses: agent guardrails, reranking, visual inspection and voice-command routing.