Liquid AI releases open d1 decision models
Liquid AI has released d1-3B and d1-omni-600M as open-weight models that return an answer in one forward pass rather than generating tokens, targeting real-time decisions on edge hardware.
Liquid AI announced two open-weight models in its d1 "decision model" family, d1-3B and d1-omni-600M. Both are on Hugging Face now, and llama.cpp supports them from launch.
Liquid's generative LFMs produce output token by token. The d1 models don't generate tokens at all: each one returns its answer in a single forward pass. According to the company, d1-3B scores 48.57 on the public split of the Decision Index v0.2.1. Liquid says that beats every model under 10B and matches Decider 35B-A3B, a model 12 times its size. Across seven public text benchmarks, d1-3B averages 82.9, compared with 81.1 for Decider 4B.
d1-3B is built on Liquid's decoder-only LFM2.5-VL-3B and accepts text and images. Liquid says that for this model, training on long inputs, shuffling answer options and fixing shortcuts in the data mattered more than more advanced techniques.
d1-omni-600M is described as experimental. It is built on a bidirectional encoder and takes either text and image or text and audio as input. It scores 15.95 on the Decision Index and averages 78.4 on the text benchmarks, ahead of Decider 2B's 77.1.
Liquid reports end-to-end latency only for d1-3B. A single question takes 8 ms on an RTX 4090, 16 ms on a Jetson AGX Thor, 26 ms on a Jetson AGX Orin and 50 ms on a Jetson Orin Nano. Answering three questions about the same input takes about 1.3 times as long as answering one.
Some results are missing from the release. Liquid did not publish results on the private vision split of Decision Index v0.3. It also says there are no dedicated benchmarks yet for audio decision tasks.