Liquid AI Opens Up d1 Decision Models for Real-Time, Edge AI Choices

Liquid AI has released two open-weight models, d1-3B and d1-omni-600M, as part of its new d1 decision model family ¹. Unlike standard language models that generate text word by word, these models read text, images, or audio and return a set of typed, probability-calculated answers in a single step — with zero output tokens generated ⁴. The goal is real-time decision-making on NVIDIA hardware, including DGX servers, RTX workstations, and Jetson edge devices .

Both models are available on Hugging Face and support the Transformers library and llama.cpp from day one ¹. The LFM Open License v1.0 allows free commercial use for companies earning under $10 million annually. d1-3B handles text and images, while d1-omni-600M can process text with either an image or audio — but not both at once .

How Decision Models Work

A generative large language model writes its answer token by token, and your code has to parse that output. A decision model skips that entirely ¹. It takes a situation and a set of named questions, reads them once, and returns a probability for every allowed answer .

There are three question types: a Noul for yes-or-no answers, a Choice for picking one label from a list, and a Score for rating something on an ordered scale from 2 to 10 levels ². Several questions about the same situation can be asked in one call, and every response reports zero output tokens ¹.

Liquid AI recommends these models for routing support tickets, content moderation, intent classification, scoring, and acting as guardrails for AI agents ¹. They are not chat models — they do not write text .

Performance and Speed

On the Decision Index benchmark, d1-3B scored 48.57, beating every model under 10 billion parameters and just behind Winnow-12B at 50.02 ¹. On seven public text benchmarks, it averaged 82.9, ahead of Decider 4B at 81.1 . d1-omni-600M, still an early research release with no published latency figures, averaged 78.4 across the same tests .

Speed is where these models shine. On an RTX 4090, one question takes just 8 milliseconds ¹. On a Jetson AGX Thor edge board, it takes 16 milliseconds — and three questions over the same state take only 20 milliseconds . That makes them suitable for live camera feeds and real-time voice routing on small devices .

Building for the Edge

d1-3B is built on the LFM2.5-VL-3B vision-language model, while d1-omni-600M starts from the LFM2.5-Encoder-350M backbone, which includes separate encoders for images and audio ¹. Audio support covers English speech up to 30 seconds .

Liquid AI also showcased d1-3B running live camera moderation and gesture control through its Open d1 Arcade demo, and navigating simulations on a Jetson device ¹. The NVIDIA Jetson AI Lab has published setup guides for the model .

Key facts

• d1-3B and d1-omni-600M are open-weight decision models that return typed answers with zero output tokens
• d1-3B scores 48.57 on Decision Index v0.2.1, beating every model under 10B parameters
• One question takes 8 ms on an RTX 4090 and 16 ms on a Jetson AGX Thor
• d1-omni-600M supports text with images or audio, but not both at once
• The LFM Open License v1.0 allows free commercial use below $10 million annual revenue

Seen on social

Announcing d1, our first decision model. It's the first model to outperform Jev on @huggingface 's Decision Index. > wins on multilingual evals > more robust against prompt injection > handles longer inputs more effectively > built for fast, structured decision-making in… pic.twitter.com/sELDdh65Ve

— Liquid AI on X, see the post

Sources

1. Liquid AI Releases Open-Weight d1-3B and d1-omni-600M: Multimodal Decision Models With Zero Output Tokens — marktechpost.com
2. Decision Models - Liquid Docs
3. Introducing d1: The most capable decision model, now with vision
4. LiquidAI releases d1-3B and d1-omni-600M decision models with zero ... — korshunov.ai
5. Multimodal open d1 decision models for the edge - Hugging Face