Liquid AI releases open decision models for edge hardware, d1-3B leads sub-10B Decision Index

Liquid AI has released two open decision models built to run on edge hardware: d1-3B and d1-omni-600M. Unlike generative models, decision models do not produce tokens; they return a structured answer in a single forward pass. d1-3B scores 48.57 on Decision Index v0.2.1, the highest result in the under-10-billion-parameter category, outperforming every 4B and 9B model as well as Decider 35B-A3B at 47.11.

Dual architecture on LFMs foundations

Both models are built on the Liquid Foundation Model family but come from entirely separate branches. d1-3B derives from LFM2.5-VL-3B, a decoder-only vision-language model (VLM) that accepts text and images. d1-omni-600M is built on LFM2.5-Encoder-350M, a bidirectional encoder fused with vision and audio encoders, enabling it to handle text-image or text-audio pairs. The smaller model is designated an early research release and remains under active development.

Benchmarks: decision quality versus size

Across seven public datasets covering reading comprehension, toxicity detection, intent classification, medical question answering and multilingual understanding, d1-3B averages 82.9, topping the table and exceeding Decider 4B. d1-omni-600M reaches 78.4, surpassing Decider 2B at 77.1 with one-quarter the parameters. Liquid notes that d1-3B's vision capabilities were preserved on standard vision benchmarks, but it does not report dedicated vision or audio scores: Decision Index v0.3 includes only a private vision split, and audio decision benchmarks are still considered an open problem.

Edge and GPU performance with NVIDIA

In collaboration with NVIDIA, d1-3B was tested on four platforms: GeForce RTX 4090, Jetson AGX Thor, Jetson AGX Orin 64 GB, and Jetson Orin Nano. Single-query latency stays below 50 milliseconds on every device — 16 ms on AGX Thor, 26 ms on AGX Orin, 50 ms on Orin Nano. Batch efficiency is notable: three concurrent queries require only 1.3× the single-query time (on Thor, 16 ms rises to 20 ms). On GPU, latency drops under 10 ms, and 384-pixel image processing completes in under 18 ms on both platforms.

Limitations and practical use

d1-omni-600M has not yet received speed figures because it remains an early research release. The models ship with custom code and require loading with `trust_remote_code=True` and transformers 5.14 or later. The API supports multiple query types on the same textual state in a single pass: yes/no (bool), multiple choice with criteria (choice), and numerical scoring (score) — for example, routing an inquiry to the appropriate team and assessing urgency simultaneously. An image can also serve as the complete input state.

Bottom line: structured decisions on device

The proposition is clear: when fast, structured decisions with multimodal input are needed on edge hardware, d1-3B delivers the highest decision quality in its size class, and d1-omni-600M fits where footprint is the binding constraint. The gap between a decision model and a generative model is not merely semantic; a single forward pass without token decoding changes the latency economics at the edge.