Liquid AI Releases d1-3B and d1-omni-600M Edge Models

Liquid AI has released two open-weight decision models, d1-3B and the experimental d1-omni-600M, designed specifically for edge computing. Unlike traditional generative models that output text token by token, these decision models process inputs and deliver structured answers in a single forward pass. This architectural choice makes them highly efficient for applications requiring rapid, real-time classification and decision-making.

The larger d1-3B model is trained on the LFM2.5-VL-3B backbone and supports text and image inputs. On the Decision Index 0.2.1, d1-3B achieved a score of 48.57, making it the top-performing model under 10 billion parameters and outperforming the much larger Decider 35B-A3B, which scored 47.11. Across seven public datasets covering tasks like medical QA and toxicity detection, d1-3B reached a mean score of 82.9. Meanwhile, the smaller d1-omni-600M model, built on the LFM2.5-Encoder-350M backbone, supports text, images, and audio. It achieved a mean score of 78.4 on the same benchmarks, beating the Decider 2B model despite having only a quarter of its parameters.

In terms of speed, d1-3B exhibits remarkable efficiency on edge hardware. Tested on the NVIDIA stack, the model answers a single query in 16 milliseconds on a Jetson AGX Thor, 26 milliseconds on a Jetson AGX Orin, and 50 milliseconds on a Jetson Orin Nano. When processing multiple queries, the model scales efficiently; answering three questions on the AGX Thor takes only 20 milliseconds. On standard GPU platforms like the NVIDIA GeForce RTX 4090, d1-3B resolves questions in under 10 milliseconds and processes a 384-pixel image in under 18 milliseconds.

For practitioners, these models offer a powerful alternative for low-latency, structured tasks on local devices. Instead of deploying heavy generative pipelines, developers can use d1-3B for high-quality multimodal decisions or d1-omni-600M for resource-constrained environments. The models are currently available on Hugging Face and require the transformers library version 5.14 or higher to run.