Liquid AI launches d1-3B and d1-omni-600M for edge computing

Liquid AI has launched two open-weight decision models targeting rapid decisions on NVIDIA devices. The announced latencies fall below 50 ms on Jetson and under 10 ms on GPU, with caveats regarding vision and audio benchmarks and the absence of speed figures for the omni variant in research.

Measured Latencies and Scope of Published Figures

According to the published measurements, d1-3B responds in less than 50 ms to a question on each tested edge device, and three questions only take 1.3 times the time of a single one, with a transition from 16 ms to 20 ms on AGX Thor. In collaboration with NVIDIA, the evaluation focused on RTX 4090, Jetson AGX Thor, Jetson AGX Orin 64 GB, and Jetson Orin Nano. On GPU, d1-3B remains under 10 ms for a question and processes a 384px image in less than 18 ms, with a state of 3.4K tokens and 64 packed states. However, Liquid AI does not include vision or audio benchmarks, with Decision Index v0.3 featuring a private vision separation and audio benchmarks remaining an open issue. No speed figures are provided for d1-omni-600M in this version due to its early research status.

Published Scores and Positioning of the Two Models

d1-3B boasts a score of 48.57 on the Decision Index 0.2.1, surpassing 4B and 9B models as well as Decider 35B-A3B, which is credited with 47.11. Both models were evaluated on seven public datasets covering reading comprehension, toxicity, intent, medical, and interlingual tasks. d1-3B claims an average score of 82.9, presented as the highest on the leaderboard and above Decider 4B. d1-omni-600M is announced at 78.4, higher than Decider 2B at 77.1, with only a quarter of the parameters reported for the latter.

Announced Architecture and Modalities

The d1 decision models rely on Liquid Foundation Models and respond in a single forward pass, without token generation. d1-3B is derived from LFM2.5-VL-3B, a visual-only decoder model, and accepts text and images; Liquid AI indicates that it retains the vision capabilities of this backbone on standard benchmarks. d1-omni-600M is based on LFM2.5-Encoder-350M, a bidirectional encoder, to which encoders for vision and audio are added to support three modalities, with text+image or text+audio inputs; according to Liquid AI, it is capable of processing these three modalities. This version is described as an initial research stage and development.

Edge and GPU Execution: Testing Conditions

On Jetson AGX Thor, d1-3B responds in 16 ms; on Jetson AGX Orin, in 26 ms; and on Jetson Orin Nano, in 50 ms. The edge measurements mention a state of 3.4K tokens and 64 packed states. On GPU, the model responds to a question in less than 10 ms and processes a 384px image in less than 18 ms, with the same reported state parameters.

Availability, Setup, and Targeted Uses

Liquid AI releases d1-3B and d1-omni-600M as open-weight models. The publisher recommends these models for obtaining rapid and structured decisions, including with multimodal inputs, presenting d1-3B as the best quality for its size and d1-omni-600M as suitable when footprint matters. Installation requires transformers version 5.14 or higher, and loading is done with trust_remote_code enabled. Both models are available on Hugging Face, with demos via System One Arcade. A blog citation reference is provided by Liquid AI with the year 2026.