LiquidAI d1-3B and d1-omni-600M Open Decision Models for Edge Inference

TL;DR

LiquidAI announced two open‑weight multimodal decision models, d1-3B and the experimental d1-omni-600M, which achieve the highest Decision Index scores among sub‑10B models and run inference in under 50 ms on a range of NVIDIA Jetson edge devices.

Why d1‑3B and d1‑omni‑600M matter

  • Decision‑oriented architecture – Unlike generative LLMs, these models produce a single structured answer per forward pass, enabling deterministic, low‑latency decisions.
  • Edge‑ready performance – d1‑3B answers a question in 16 ms on a Jetson AGX Thor and under 50 ms on all tested Jetson platforms, making real‑time on‑device inference feasible.
  • Multimodal flexibility – d1‑3B handles text + image inputs; d1‑omni‑600M extends this to text + image or text + audio, supporting a broader range of sensor data.

Model architecture and training

d1‑3B

  • Built on LFM2.5‑VL‑3B, a decoder‑only vision‑language foundation model.
  • Accepts text and 384 px images as inputs.
  • Trained as a decision model that emits a single answer vector rather than token sequences.

d1‑omni‑600M (experimental)

  • Derived from LFM2.5‑Encoder‑350M, a bidirectional encoder.
  • Adds separate vision and audio encoders, enabling two modality configurations: text + image or text + audio.
  • Currently a research release; speed numbers are not reported.

Benchmark performance

The models were evaluated on seven public datasets covering reading comprehension, toxicity detection, intent classification, medical QA, and cross‑lingual understanding. Scores are reported on the Decision Index v0.2.1.

Benchmark d1‑omni‑600M d1‑3B Decider 2B Decider 4B
SQuAD 2.0 74.0 83.3 67.7 76.0
Civil Comments 95.8 93.3 93.6 92.8
MASSIVE intent 86.1 86.9 81.1 88.3
PubMedQA 61.3 68.3 65.7 63.3
BoolQ 77.7 86.3 87.3 89.0
XNLI 74.7 85.6 85.0 88.6
PAWS‑X 79.5 76.4 59.5 69.8
Mean 78.4 82.9 77.1 81.1
  • d1‑3B achieves the highest mean score (82.9) among all listed models, surpassing the 4 B Decider baseline.
  • d1‑omni‑600M reaches 78.4, outperforming Decider 2B despite having only 600 M parameters.
  • Vision capabilities of d1‑3B were verified on standard vision benchmarks; audio benchmarks for d1‑omni‑600M are not yet released.

Edge and GPU latency

Edge devices (NVIDIA Jetson)

Device One question Three questions 3.4K‑token state 384 px image 64 packed states
Apple M5 Pro 30 ms 41 ms 640 ms 62 ms 78 /s
Jetson AGX Thor 16 ms 20 ms 220 ms 35 ms 262 /s
Jetson AGX Orin 64 GB 26 ms 35 ms 560 ms 83 ms 110 /s
Jetson Orin Nano 50 ms 73 ms 1,640 ms 202 ms 38 /s
  • Latency scales sub‑linearly: processing three questions costs only ~1.3× the time of a single query on the Thor.

GPU platforms

Platform One question Three questions 3.4K‑token state 384 px image 64 packed states
NVIDIA RTX 4090 8 ms 21 ms 102 ms 17 ms 475 /s
AMD MI325X 9 ms 14 ms 44 ms 18 ms 1,106 /s
  • On high‑end GPUs, d1‑3B answers a question in under 10 ms and processes a 384 px image in under 18 ms.

Getting started

The models are distributed via Hugging Face with self‑contained inference code. Example code for d1‑3B (requires transformers>=5.14):

pip install "transformers>=5.14" torch torchvision pillow

import io, urllib.request
import torch
from PIL import Image
from transformers import AutoModel

device = "cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu"
model = AutoModel.from_pretrained(
    "LiquidAI/d1-3B",
    trust_remote_code=True,
    dtype=torch.float32 if device == "cpu" else torch.bfloat16,
).to(device)

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 unauthorised use"}},
    "urgency": {"type": "score", "instructions": "How urgent is this?",
                "criteria": ["Can wait", "Today", "Blocking the customer now"]},
}
print(model.system_one("I was charged twice this month, please refund one of them.", questions))

# Image example
url = "http://images.cocodataset.org/val2017/000000039769.jpg"
photo = Image.open(io.BytesIO(urllib.request.urlopen(url).read()))
print(model.system_one(None, {"cats": {"type": "choice", "instructions": "How many cats are there?",
               "criteria": {"one": "One", "two": "Two", "more": "Three or more"}}}, images=[photo]))

# Batch example
tickets = ["Where is my parcel? It was due Monday.", "The app crashes when I open settings."]
print(model.system_one_batch([(t, {"team": questions["team"]}) for t in tickets]))
  • The d1‑omni‑600M card provides analogous usage instructions for text + image or text + audio inputs.

Availability and community resources

Implications for edge AI

  • Deterministic, low‑latency decisions enable on‑device moderation, fraud detection, and real‑time assistance without round‑trip cloud latency.
  • Multimodal input broadens applicability to robotics, IoT, and mobile assistants that must fuse visual or acoustic cues with textual context.
  • Open‑weight release lowers the barrier for developers to experiment, benchmark, and integrate decision models into proprietary pipelines, potentially accelerating adoption of edge‑centric AI.

LiquidAI’s open d1 decision models set a new performance‑to‑size baseline for multimodal edge inference, offering both high accuracy and sub‑50 ms latency on commodity Jetson hardware.

Sources