[Models](/models)/LiquidAI/d1-3B

LFM Open License v1.0 (commercial use under $10M annual revenue)

# d1-3B

by LiquidAI

A 3B open decision model that answers yes/no, choice and rating questions over text and images in one pass, 8 ms on an RTX 4090

104likes

3.1Bparams

[HuggingFace](https://huggingface.co/LiquidAI/d1-3B)[Run it on your own data, free](https://studio.mixpeek.com/namespaces/create?mode=managed&utm_source=homepage&utm_medium=model_page&utm_campaign=mvs_onramp)

Identifiers

Model ID

`LiquidAI/d1-3B`

## Deploy d1-3B

Single-tenant

Mixpeek has no managed extractor for this model. On a single-tenant deployment you upload the weights and a custom plugin serves them next to the rest of your pipeline.

[Read the customer-hosted guide](https://mixpeek.com/docs/customer-hosted/overview) [Talk to us about single-tenant](/demo)

## Overview

d1-3B is Liquid AI's open decision model, released in October 2026 and built on LFM2.5-VL-3B. Give it a state, which can be text, JSON, images or a mix, and a set of named questions, and it returns a calibrated probability for each answer in one forward pass with no generated text.

Liquid reports it as the best decision model under 10B on the Decision Index 0.2.1 (48.57, ahead of a 36B model at 47.11) and fast enough for real-time pipelines: 8 ms a question on an RTX 4090 and 30 ms on an Apple M5 Pro. The weights are free for commercial use only for organisations under $10M in annual revenue.

## Architecture

LFM2.5-VL-3B (3.12B parameters, SigLIP2 NaFlex vision encoder, 32,768-token context, 128K vocabulary) post-trained to read each answer directly from the model's distribution over the options. Questions follow the Decision Index schema: a type (noul for yes/no, choice, or score), instructions and criteria. The state and images are read once for all questions in a call.

## Mixpeek SDK Integration

IngestRetrieve

\# Decision models judge candidates; index and retrieve the content with Mixpeek first.
import requests

docs = requests.post(
    "https://api.mixpeek.com/v1/retrievers/ret\_your\_retriever/execute",
    headers={"Authorization": "Bearer API\_KEY", "X-Namespace": "ns\_your\_namespace"},
    json={"inputs": {"query": "damaged packaging on arrival"}},
).json()\["documents"\]

## Capabilities

*   Yes/no, pick-one and 2-to-10-level rating questions, answered as probabilities with zero output tokens
*   Several named questions over one state in a single pass
*   Text, JSON and images in the same state; 32,768-token context
*   Batching that packs many states into one pass with no padding
*   Runs on data-centre GPUs, RTX workstations, Apple silicon and Jetson

## Use Cases on Mixpeek

Filtering or reranking retrieved results with a probability you can threshold

Moderation and visual inspection of images at the edge

Routing support tickets and classifying intent in real time

Agent guardrails and LLM-as-a-judge scoring without generating text

## Benchmarks

Dataset

Metric

Score

Source

Decision Index 0.2.1

Score

48.57

Model card: LiquidAI/d1-3B (Liquid ran the official scorer itself, without a leaderboard submission; Decider 35B-A3B 47.11)

8 public benchmarks as decisions (SQuAD 2.0, BoolQ, XNLI and others)

Mean

77.1

Model card (Liquid's internal evaluation; Decider 4B 76.2)

11 public image benchmarks as decisions

Mean

74.1

Model card (base LFM2.5-VL-3B 73.9)

DecisionBench (eng v1, 23,900 rows)

Score

71.8

Model card (self-reported)

One question, NVIDIA RTX 4090, bf16

Latency

8 ms

Model card (warm, with model.compile; 16 ms without)

## Performance

Input SizeText, JSON and images; 32,768-token context

Embedding Dimn/a (returns a probability per option)

GPU LatencyInput dependent

GPU ThroughputBatch dependent

GPU MemoryModel dependent

Liquid's figures: 8 ms a question on an RTX 4090, 9 ms on an AMD MI325X, 30 ms on an Apple M5 Pro, 50 ms on a Jetson Orin Nano; 475 states a second packed on the 4090. The first call with a new shape pays for compilation, so warm up. We have not measured it.

## Common Pipeline Companions

[

autotrust/JEV-27B-VL

A larger open decision model over text and images, Apache-2.0

](/model/autotrust/JEV-27B-VL)

## Frequently Asked Questions

### What is d1-3B?

An open decision model from Liquid AI. It answers questions about text or images, such as yes or no, which of these options, or how severe on a scale, with a probability for each answer and no generated text.

### How fast is d1-3B?

Liquid reports 8 ms for one question on an NVIDIA RTX 4090, 9 ms on an AMD MI325X and 30 ms on an Apple M5 Pro, warm and one request at a time. Packing 64 states into one pass reaches 475 a second on the 4090.

### Can I use d1-3B commercially?

Under the LFM Open License v1.0, commercial use is allowed for organisations with less than $10M in annual revenue. Above that you need an agreement with Liquid AI.

### How do I use d1-3B with Mixpeek?

Run it on the candidates a Mixpeek retriever returns, to keep, label or reorder them with a probability per item, as in the example on this page. Mixpeek's built-in decision stages call TypeSafe AI's hosted Jev by default.

## Explore on Mixpeek

[

Browse all models

Search the full model catalog

](/models)[

Feature Extractors

See how Mixpeek runs models as extractors

](/extractors)[

Mixpeek Vector Store

Store & search embeddings at scale

](/mvs)[

Pricing

Usage-based pricing for pipelines

](/pricing)[

All Curated Lists

Compare models, APIs & infrastructure

](/curated-lists)

### Specification

OrganizationLiquidAI

Retriever\-

Parameters3.1B

LicenseLFM Open License v1.0 (commercial use under $10M annual revenue)

Downloads/moN/A

Likes104

[

### Research Paper

Open d1: Edge decision models for text, vision, and audio (Liquid AI)

arxiv.org](https://www.liquid.ai/blog/d1-open)

### Build a pipeline with d1-3B

Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.

[Run it on your own data, free](https://studio.mixpeek.com/namespaces/create?mode=managed&utm_source=homepage&utm_medium=model_page&utm_campaign=mvs_onramp)