Liquid AI 发布 Open d1：零输出 Token 的开放权重决策模型

RT Marktechpost AI Liquid AI just released Open d1, 2 open-weight decision models that never generate a single token. d1-3B (text + vision) and d1-omni-600M (text + image, or text + audio) read a state plus typed questions and return calibrated probabilities in 1 forward pass. The core idea is to replace "generate JSON, then parse it" with a direct readout. Every question is a noul (P(yes)), a choice (a distribution over named options) or a score (a position on a 2 to 10 level rubric). Output tokens: always 0. d1-3B scores 48.57 on Decision Index v0.2.1. That is the best result under 10B, and above Decider 35B-A3B (47.11), a model 12x its size. The trade-off: it trails on Knowledge (23.8 vs 33.8 for Winnow-12B). It is a decision engine, not a chat model. Latency per question: 8 ms on RTX 4090, 16 ms on Jetson AGX Thor, 50 ms on Jetson Orin Nano. Day-one llama.cpp support...... Full analysis: https://www.marktechpost.com/2026/10/07/liquid-ai-releases-open-weight-d1-3b-and-d1-omni-600m-multimodal-decision-models-with-zero-output-tokens/ d1-3B: http://huggingface.co/LiquidAI/d1-3B d1-omni-600M: http://huggingface.co/LiquidAI/d1-omni-600M Docs: http://docs.liquid.ai/lfm/models/decision-models @liquidai @nvidia