What are Liquid AI's new d1 models and why do they answer in one pass?

Decision models do not generate text

A decision model takes a question and returns an answer in a single forward pass. There is no token-by-token generation, no streaming, no chain of thought. The output is a structured object — yes or no, a choice from a list, or a numeric score — produced all at once. They answer instantly. Liquid AI built d1-3B and d1-omni-600M on top of their Liquid Foundation Models but changed the training objective so the network learns to classify instead of continue.

Two architectures, two modality sets

d1-3B starts from LFM2.5-VL-3B, a decoder-only vision-language model. It accepts text and images. d1-omni-600M starts from LFM2.5-Encoder-350M, a bidirectional encoder, and adds separate vision and audio encoders. It accepts either text with an image or text with audio, but not all three at once. The smaller model is labeled an early research release and is still undergoing development.

Benchmarks across seven public datasets

The team evaluated both models on SQuAD 2.0, Civil Comments, MASSIVE intent, PubMedQA, BoolQ, XNLI, and PAWS-X. d1-3B reaches a mean score of 82.9, the highest in the published table and above Decider 4B at 81.1. d1-omni-600M scores 78.4, edging out Decider 2B at 77.1 while using roughly a quarter of the parameters. On the Decision Index 0.2.1 d1-3B scores 48.57, ahead of every 4B and 9B model and ahead of Decider 35B-A3B at 47.11. Vision and audio benchmarks are not reported because the Decision Index v0.3 contains only a private vision split and audio decision benchmarks remain an open problem.

Speed on edge hardware

NVIDIA collaborated on measurements across four devices. d1-3B answers one question in 16 ms on a Jetson AGX Thor, 26 ms on a Jetson AGX Orin 64 GB, and 50 ms on a Jetson Orin Nano. Three questions take only 1.3× the time of one; on the AGX Thor the latency rises from 16 ms to 20 ms. On GPU the same model answers in 8 ms on an RTX 4090 and 9 ms on an AMD MI325X. A 384-pixel image adds 17 ms on the 4090 and 18 ms on the MI325X. No speed numbers are published for d1-omni-600M.

Running the models locally

Install transformers 5.14 or newer, torch, torchvision, and pillow. Load the model with trust_remote_code=True because the repository ships its own modeling code. The system_one method accepts a state — text, an image, or both — and a dictionary of named questions. Each question defines a type (noul, choice, or score), instructions, and criteria. A batch variant system_one_batch processes multiple states with the same question set in one call. The d1-omni-600M model card contains separate usage instructions.

What the release does not cover

The blog post does not disclose training data composition, licensing details for the weights, or power consumption figures for the Jetson modules. It also does not compare latency on Apple silicon beyond the single M5 Pro column shown in the speed table. Audio decision quality for d1-omni-600M has no public benchmark yet.