Liquid AI’s d1-3B makes decisions in one pass

Liquid AI has released d1-3B, an open-weight multimodal model that returns structured predictions in a single forward pass and generates zero output tokens. Built on Liquid Foundation Models, it removes the sequential decoding loop used by generative transformers. Pipelines that need a yes-or-no answer, category, ranking, or score can therefore avoid generating and parsing text.

The 3.1B-parameter model builds on LFM2.5-VL-3B. Liquid also released an experimental 600M sibling called d1-omni-600M. Both models are available on Hugging Face.

One pass, three answer types

Each request contains a state, such as text, JSON, an image, or a combination, plus a dictionary of named questions. The model reads the state once and returns typed answers with probabilities that Liquid describes as calibrated. The API defines three primitives:

• Noul: Answers a yes-or-no question with a probability from 0 to 1. For example, “Is this message spam?” might return 0.92.
• Choice: Selects one named option and returns a probability distribution across all supplied options.
• Score: Rates the input against an ordered rubric and returns a probability-weighted position on that scale.

A single API call can apply several decisions to the same customer message:

questions = {
    "refund": {
        "type": "noul",
        "instructions": "Is the customer asking for a refund?",
    },
    "team": {
        "type": "choice",
        "instructions": "Which team should handle this?",
        "criteria": {
            "billing": "Charges, refunds, invoices",
            "technical": "App or site faults",
            "fraud": "Suspected unauthorized use",
        },
    },
    "urgency": {
        "type": "score",
        "instructions": "How urgent is this?",
        "criteria": [
            "Can wait",
            "Today",
            "Blocking the customer now",
        ],
    },
}

model.system_one(
    "I was charged twice this month, please refund one of them.",
    questions,
)