Liquid AI's d1-3B Decides in 8ms: Open-Weight 3B Decision Model Released

Introduction

Liquid AI has open-sourced two decision-focused models on Hugging Face: the 3B-parameter d1-3B and the 600M-parameter multimodal d1-omni-600M. Unlike conventional generative LLMs that decode token-by-token, these models perform a single forward pass to output classification, scoring, or routing decisions directly. Official benchmarks show a single decision completes in 8 milliseconds on an RTX 4090 (with CUDA Graph compilation), and the output token count is zero, eliminating hallucination and format-failure risks.

Why Decision Models?

Most deployed LLMs are generative, requiring lengthy autoregressive decoding even for simple yes/no or routing tasks. In production, many high-volume needs only require a decision:

Agent tool selection

Customer ticket routing

Compliance, toxicity, and safety filtering

RAG reranking and quality scoring

Automated visual defect grading

Running these on large generative models incurs hundreds of milliseconds of latency, heavy VRAM usage, and fragile JSON-schema constraints. The article includes a diagram comparing traditional generative LLM pipelines with the d1 single-forward-decision approach.

Architecture and Multimodal Family

1. d1-3B: Decision Ceiling in a Small Package

Backbone: Liquid's latest multimodal vision model LFM2.5-VL-3B (Decoder-only).

Training: Fused text-backbone weights from LFM2.5-2.6B and LFM2.5-VL-3B, then fine-tuned with long-input handling, option-order shuffling, and data-shortcut fixes.

Vision encoder: SigLIP2 NaFlex 400M.

Context window: 32K tokens, supporting joint text-and-image decisions.

An architecture diagram illustrates the d1-3B structure and weight-merging training flow.

2. d1-omni-600M: Edge-Ready Multimodal

Base: LFM2.5-Encoder-350M bidirectional encoder.

Audio: FastConformer encoder plus adapter, trained in stages.

Vision: Borrows LFM2.5-VL-450M encoder with LoRA updates.

Parameters: 600M total, accepts mixed text+image or text+audio, small enough for edge devices.

A second diagram shows the text-vision-audio fusion architecture.

Performance and Benchmarks

On the Decision Index v0.2.1 public benchmark, d1-3B scores 48.57, outperforming all sub-10B models and even surpassing the proprietary 35B-parameter Decider 35B-A3B (47.11). The Tools sub-score reaches 74.5, making it attractive for agent frameworks.

End-to-end latency (single decision):

RTX 4090: 8 ms (CUDA Graph), 16 ms uncompiled

AMD MI325X: 9 ms, 1,106 decisions/sec throughput

Jetson AGX Thor: 16 ms

Jetson AGX Orin 64GB: 26 ms

Jetson Orin Nano: <50 ms

Apple M5 Pro (MPS): 30 ms

State reuse: asking three decisions on the same context costs only ~1.3× a single decision, not 3×.

Hugging Face Spaces hosts 10 zero-setup demos (gesture-controlled games, real-time content moderation, etc.). Benchmark chart and demo screenshots are included.

How to Use

Weights are on Hugging Face; local inference requires transformers>=5.14.

pip install "transformers>=5.14" torch torchvision pillow

Python example loading d1-3B and defining a decision schema with the three primitives ( noul, choice, score):

import torch
from transformers import AutoModel
from transformers.image_utils import load_image

device = "cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu"
dtype = torch.float32 if device == "cpu" else torch.bfloat16

# Load d1-3B
model = AutoModel.from_pretrained(
    "LiquidAI/d1-3B",
    trust_remote_code=True,
    dtype=dtype
).to(device)

# Define decision schema
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"]
    }
}

# Single forward inference
result = model.system_one(
    "I was charged twice this month, please refund one of them.",
    questions
)

print(result)

Image input is equally simple: pass images=[image]. Day-one llama.cpp support covers Apple Silicon, AMD, Qualcomm, and NVIDIA NVFP4 quantization.

Commercial Licensing Caveat

Open-weight does not mean unrestricted commercial use. The models use Liquid AI's LFM Open License v1.0:

Annual revenue under $10M: free commercial use and derivative development for individuals, researchers, and small teams.

Annual revenue $10M or above: must contact Liquid AI for a dedicated commercial license.

Teams should evaluate both the performance gains and their compliance boundaries.

Conclusion

For developers previously locked into cloud-only, high-latency, token-costly closed models, Liquid AI's open-weight release is timely. The combination of tiny footprint, single-forward inference, millisecond response, and zero token consumption directly addresses the determinism and concurrency pain points in agent systems. While the 8 ms figure reflects a best-case CUDA Graph run and real workloads add preprocessing overhead, the leap over multi-second generative model latency is qualitative. Interested practitioners can pull the weights from Hugging Face immediately.