Liquid AI's d1 models answer without writing a single token
d1-3B and d1-omni-600M are open-weight "decision models" that skip text generation entirely, returning a calibrated answer in one forward pass in as little as 8ms.
Liquid AI released d1-3B and d1-omni-600M on 7 October, open-weight models built to answer a typed question — a choice, a score, a yes/no — without generating any tokens at all. d1-3B is trained from Liquid’s LFM2.5-VL-3B vision-language backbone and takes text and images; d1-omni-600M is built on a bidirectional encoder with added vision and audio, so it can also answer on sound.
On Liquid’s own Decision Index v0.2.1 benchmark, d1-3B scores 48.57 overall, with a mean of 82.9 across its seven component benchmarks. The pitch is latency: a single question answers in 8ms on an RTX 4090, 16ms on an NVIDIA Jetson AGX Thor, and 30ms on an Apple M5 Pro. Both models are on Hugging Face with day-one support in llama.cpp.
# a decision model returns an answer, not tokens
output_tokens: 0Why it matters: if a pipeline calls a full generative model just to triage a ticket, score a response, or classify an image, that’s a decision, not a generation task — and a model built for it removes the sampling step and the latency that comes with it, on hardware as small as a Jetson.
The caveat: these aren’t Apache or MIT. Both ship under the LFM Open License v1.0, which is free for commercial use only below $10 million in annual company revenue; past that, Liquid requires a separate paid licence.