Two open checkpoints, no tokens
Liquid AI released two open-weight models on 7 October: d1-3B, which takes text and images, and d1-omni-600M, which takes text with either an image or audio. Both are on Hugging Face, and both belong to a category that has been filling up fast this autumn — decision models, which do not generate text at all.
The distinction matters for latency. A generative model emits tokens one at a time; a decision model processes the state in a single forward pass and returns a structured answer from options the developer supplies. Liquid AI therefore measures end-to-end time rather than tokens per second, and the numbers it publishes are in milliseconds.
What the company reports
On Liquid AI’s own Decision Index v0.2.1, public split, the company puts d1-3B at 48.57 — ahead, it says, of every model under 10B parameters and on a par with Decider 35B-A3B, a model twelve times its size. d1-omni-600M scores 15.95 on the same index and is described as a first experimental checkpoint.
Across seven public text benchmarks — SQuAD 2.0, Civil Comments, MASSIVE intent, PubMedQA, BoolQ, XNLI and PAWS-X — the company reports a mean of 82.9 for d1-3B against 81.1 for Decider 4B, and 78.4 for d1-omni-600M against 77.1 for Decider 2B at roughly a quarter of the parameters. The small model leads the table on toxicity detection, at 95.8 on Civil Comments, and on paraphrase identification, at 79.5 on PAWS-X. Every one of these figures is Liquid AI’s own; none has been independently reproduced.
The company is explicit about one gap. Decision Index v0.3 includes a private vision split, and Liquid AI says it is not reporting on it in this release, offering instead the observation that d1-3B retains the vision ability of its LFM2.5-VL-3B backbone on standard benchmarks. For audio, it says dedicated decision benchmarks are “currently an open problem” and invites the community to build them.
Where the speed claim comes from
The latency table is the part worth reading closely, because it is specific. For a single question, Liquid AI measures d1-3B at 8ms on an Nvidia RTX 4090 and 9ms on an AMD MI325X; at the edge, 16ms on an Nvidia Jetson AGX Thor, 26ms on a Jetson AGX Orin 64GB, 30ms on an Apple M5 Pro and 50ms on a Jetson Orin Nano.
Three questions over the same state cost only about 1.3 times a single one — 16ms to 20ms on the AGX Thor — which is the practical point of the architecture: one pass over the state can answer several questions about it. A 384-pixel image takes 17ms on the 4090 and 35ms on the Thor. A 3,400-token state is where the small devices struggle: 1,640ms on the Orin Nano against 102ms on the 4090. All of these are single-request measurements, and the edge figures were produced in collaboration with Nvidia.
How they were built
d1-3B is trained from LFM2.5-VL-3B, a decoder-only vision-language model. Liquid AI says it averaged the weights of LFM2.5-2.6B and the text backbone of LFM2.5-VL-3B to make a better base, fine-tuned checkpoints with different seeds and data mixtures, then merged them again — and that training on long inputs, shuffling answer options and fixing shortcuts in the data mattered more than more elaborate techniques.
d1-omni-600M comes from LFM2.5-Encoder-350M, a bidirectional encoder, with a FastConformer audio encoder and a vision encoder from LFM2.5-VL-450M attached in stages.
The models ship with day-one llama.cpp support and the company describes them as downloadable, fine-tunable and deployable without restrictions. What to watch is whether an independent evaluation of the Decision Index numbers appears, and whether anyone builds the audio decision benchmark Liquid AI says does not exist.
