Liquid AI open-sources d1-3B for sub-50ms edge decisions

Liquid AI has open-sourced d1-3B, a 3.1-billion-parameter multimodal decision model designed strictly for classification and evaluation tasks, built on top of its LFM2.5-VL-3B architecture.

It is not a conversational chatbot and does not generate descriptive text. Instead of running continuous autoregressive decoding, you feed it a state made of text, JSON, or images alongside explicit questions. The model returns calibrated, typed answers in a single forward pass with zero output tokens.

What it does

Developers can plug d1-3B directly into automated pipelines to handle routing, triage, content moderation, intent classification, LLM-as-a-judge scoring, agent guardrails, and automated visual inspection.

Queries are structured around three core types. The model can return a probabilistic yes-or-no answer, select a single choice from multiple labeled candidates with full probability distributions, or score an input across an ordered scale of two to ten levels. Because responses are delivered natively as structured data, engineers do not have to write regex parsers or deal with hallucinated formats.

On the public Decision Index 0.2.1 benchmark, d1-3B scored 48.57. That places it at the top of the under-10B category, outperforming every 4B and 9B competitor as well as the 36B Decider 35B-A3B. Across 11 standard image evaluations, it posted an average score of 74.1, matching the perceptual quality of full-sized vision-language base models.

Speed and hardware specs

By ditching token-by-token text generation, d1-3B delivers exceptionally low latency on both data center GPUs and edge modules.

On an NVIDIA RTX 4090 with compilation optimizations enabled, single-question latency drops to just 8 ms. An AMD MI325X handles it in 9 ms, while an Apple M5 Pro clocks in at 30 ms.

Edge deployments show similarly impressive figures. The model executes in 16 ms on an NVIDIA Jetson AGX Thor, 26 ms on a Jetson AGX Orin, and stays at 50 ms on a lightweight Jetson Orin Nano.

The model features 3.12 billion parameters, a 32,768-token context window, and a 400M SigLIP2 NaFlex vision encoder. It requires transformers version 5.14 or later and runs out of the box with vLLM, SGLang, and Docker. Liquid AI has hosted the weights on Hugging Face alongside an Open d1 Arcade demo collection for quick hands-on testing.

Why it matters

If you are building multi-agent workflows, customer support routing, or edge vision inspection, d1-3B is an easy recommendation. Using oversized conversational LLMs just to parse a yes-or-no decision is expensive, slow, and prone to formatting failures; d1-3B collapses that entire step into a millisecond-level forward pass.

Decoupling pure decision-making from conversational generation is a smart architectural move. We believe small, single-pass decision engines like this are far more practical for local devices and robotics than running general-purpose conversational models on the edge.