d1-3B Vision-Language Model: 4GB+ VRAM, GGUF Builds

d1-3B Vision-Language Model: 4GB+ VRAM, GGUF Builds

At a Glance

Item Value
Repository LiquidAI/d1-3B
Family guide LFM2.5 guide (11 articles)
Publisher guide Liquid AI: models and licenses
Published 2026-10-05
License lfm1.0
Formats safetensors
Paper arXiv:2511.23404
Source type Primary source (the publisher itself)

Values determined by this site’s code when the information was collected. Dates are JST.

Specifications

  • Parameters: 3.12B
  • Architecture: lfm2_vl (LFM2.5-2.6B + SigLIP2 NaFlex shape-optimized 400M)
  • Context length: 32,768 tokens
  • Vocabulary size: 128,000

Performance

According to the model card, d1-3B demonstrates performance in decision-making tasks that rivals or surpasses models significantly larger in size. Below is a comparison table of “Decision Index 0.2.1" based on measurements by the publishers. Note that open-weight models of the 10B class and below are primarily included for comparison.

→ Scroll horizontally to see all columns

Model Size Decision Index Knowledge Language Retrieval Tools Arts
Winnow-12B 12B 50.02 33.8 56.0 54.0 71.0 30.0
d1-3B 3B 48.57 23.8 56.4 52.8 74.5 36.3
Decider 35B-A3B 36B 47.11 31.8 55.5 54.7 56.5 32.6
JPT-9B 9.7B 46.89 31.7 56.7 44.6 67.0 28.6
JPT-4B 4.7B 43.04 28.7 52.5 45.0 57.2 25.8
Decider 4B 4.7B 40.70 25.7 46.0 44.7 58.6 25.0
Decider 2B 2.3B 28.97 14.9 32.6 37.3 42.4 14.6

This table shows that despite its small parameter count of 3B, d1-3B outperforms the 35B-class Decider 35B-A3B in the overall score (Decision Index). It records the highest figures among the comparison targets particularly in the “Tools" and “Arts" categories, making it exceptionally strong for agentic tasks. On the other hand, the “Knowledge" score remains at 23.8, indicating a tendency to fall short in tasks requiring broad knowledge when compared to models like the 12B-class Winnow-12B (33.8).

Next, the results of evaluating existing benchmarks as decision-making tasks are shown below.

Benchmark d1-3B Decider 4B Decider 2B
SQuAD 2.0 85.3 76.0 67.7
Civil Comments 93.0 92.8 93.6
MASSIVE intent 87.3 88.3 81.1
HelpSteer2 36.7 42.0 32.0
PubMedQA 66.0 63.3 65.7
BoolQ 86.7 89.0 87.3
XNLI 85.0 88.6 85.0
PAWS-X 76.9 69.8 59.5
Mean 77.1 76.2 71.5

In this evaluation, it shows accuracy that significantly outpaces Decider 4B in SQuAD 2.0 and PAWS-X. It also records a mean of 77.1, maintaining performance that exceeds 4B-class models.

Finally, here is a comparison with the base model LFM2.5-VL-3B regarding multimodal (image understanding) performance.

Benchmark d1-3B LFM2.5-VL-3B
AI2D 79.9 80.9
BLINK 59.2 58.7
CV-Bench 82.1 87.6
HallusionBench 65.3 65.0
MMBench 84.9 84.3
MME 82.1 82.4
MMStar 59.9 61.2
MMVP 77.0 73.7
POPE 88.5 90.1
VisualWebBench 71.4 78.3
VL-RewardBench 65.0 50.9
Mean 74.1 73.9

Its average image understanding score is 74.1, maintaining performance comparable to the base model (73.9) while being optimized as a decision model. Notably, VL-RewardBench sees a substantial improvement from 50.9 in the base model to 65.0. Meanwhile, there are items where it falls short of the base model, such as CV-Bench and VisualWebBench.

Strengths and Use Cases

d1-3B is a decision model specialized in receiving states (such as text or images) and query items, and returning calibrated, typed answers in a single forward pass with zero output tokens.

Because it is not a chat model that outputs free-form text, it cannot be used for text writing or general conversational purposes. However, it excels in applications requiring fast and accurate judgments within system decision-making processes or automation pipelines, such as the following:

  • Routing, Triage, and Intent/Topic Classification: Analyzing inbound customer inquiries or ticket contents to route them to the appropriate department or determine urgency.
  • Content Moderation and Guardrails: Determining whether input text or images violate terms of service, or functioning as a guardrail to ensure the safety of AI agents.
  • Data Extraction Checking and Reranking: Performing consistency checks on extracted information or reranking search results.
  • LLM-as-a-Judge Scoring: Evaluating and scoring the quality of responses generated by other LLMs.
  • Visual Inspection: Supporting multimodal visual discrimination tasks such as counting objects included in an input image or identifying their states.

Question formats that can be specified with this model include the following types:

  • noul (Yes/No judgment): Returns the probability of affirmation for a question.
  • choice (Multiple-choice judgment): Selects one from defined named choices and returns it along with a confidence score or probability distribution.
  • score (Scoring): Returns an expected score level and confidence from among 2 to 10 ordered levels.

Additionally, supported languages include Japanese (ja), English (en), Chinese (zh), French (fr), German (de), Hindi (hi), Indonesian (id), Italian (it), Korean (ko), Polish (pl), Portuguese (pt), Russian (ru), Spanish (es), Thai (th), Vietnamese (vi), Arabic (ar), and many others. Even when multiple questions are provided simultaneously for a single state, judgments can be obtained collectively in a single execution.

Hardware Requirements

Estimated requirements (calculated by Local Model Watch) — 3.1B parameters

Your VRAM Quantization File size Est. memory needed
4GB (laptop iGPU / phone class) Q8_0 2.7GB 3.2GB
8GB (RTX 4060 / 3060 Ti, etc.) BF16 5.0GB 6.0GB

Inference engine support (architecture name matched against each project’s own model registry in its source code, checked 2026-11-17): llama.cpp: registered, vLLM: registered, MLX (mlx-lm): not registered. “Not registered" means the name is absent from that registry today, not that the model cannot run.

Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. File sizes are measured from the converted build LiquidAI/d1-3B-GGUF. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.

Can You Run It Locally?

Runs in Ollama, LM Studio and llama.cpp via a converted build.

The publisher ships safetensors, but LiquidAI/d1-3B-GGUF provides a GGUF build you can use.

License — lfm1.0 (Commercial use allowed with conditions): LiquidAI’s LFM Open License v1.0. Commercial use is allowed, except by legal entities with annual revenue of $10 million or more (Section 5). Redistribution requires the license text and notice of changes.

Compression: the Q8_0 build measures 7.43 bits per weight — about 46% the size of the original 16-bit weights, calculated by this site from the actual file sizes.

Compiled by this site’s code from the published formats, converted builds we have found, and each engine’s own model registry. “Not found" means we have not seen such a build, not that none exists. License summaries are not legal advice — check the publisher’s original terms before relying on them.

Recent Models in the Same Size Class

Models with up to 4B parameters that Local Model Watch covered recently, listed by code from our article log for comparison. VRAM tiers are this site’s estimates; licenses are as stated on the model cards.

→ Scroll horizontally to see all columns

Model Parameters Smallest VRAM tier License Our article
LiquidAI/d1-omni-600M 587M 4GB lfm1.0 d1-omni-600M Multimodal Decision Model: 4GB+ VRAM, GGUF Builds (2026-11-17)
baidu/Unlimited-OCR-2 3.3B 8GB mit Unlimited-OCR-2 Vision-Language Model: 8GB+ VRAM (2026-11-17)
openbmb/MiniCPM-OCR-1.0-DSpark 297M 4GB apache-2.0 MiniCPM-OCR-1.0-DSpark OCR Acceleration Draft Model: 4GB+ VRAM (2026-11-17)
openbmb/MiniCPM-OCR-Jev-GPTQ 2.2B 4GB apache-2.0 OpenBMB Releases Quantized Versions of MiniCPM-OCR-Jev (2026-11-17)
LiquidAI/d1-2B-GGUF 1.9B 4GB lfm1.0 d1-2B-GGUF Text, Image, and Video Decision-Making Model: 4GB+ VRAM (2026-11-17)

How to Get It

d1-3B is distributed in safetensors format and can be downloaded from the Hugging Face Hub. Submitting a license agreement (gated) is not required for usage.

Running the model requires transformers>=5.14, torch, torchvision, and pillow as Python packages. You can install the dependencies with the following command:

pip install "transformers>=5.14" torch torchvision pillow

Because model-specific implementation code is included in the repository, specify trust_remote_code=True upon loading.

import torch
from transformers import AutoModel

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

To execute judgments, use the model.system_one() method. You can obtain answers collectively by passing a question definition dictionary and text or an image.

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

Furthermore, the model.system_one_batch() method is also provided to process multiple requests in bulk without padding, supporting batch processing for large volumes of requests.

Quantized and Converted Variants

→ Scroll horizontally to see all columns

Added Publisher Format Repository Smallest VRAM tier (build, est. memory)
2026-10-07 LiquidAI INT8 LiquidAI/d1-3B-w8a8 INT8 4.3GB (fits in 8GB VRAM)
2026-10-07 LiquidAI GGUF LiquidAI/d1-3B-GGUF Q8_0 3.2GB (fits in 4GB VRAM)
2026-10-12 LiquidAI MLX LiquidAI/d1-3B-MLX-fp16 MLX 7.0GB (fits in 8GB VRAM)
2026-10-12 LiquidAI MLX LiquidAI/d1-3B-MLX-4bit MLX 4bit 2.6GB (fits in 4GB VRAM)
2026-10-12 LiquidAI MLX LiquidAI/d1-3B-MLX-8bit MLX 8bit 4.2GB (fits in 8GB VRAM)

File sizes of each build:

  • Available builds in LiquidAI/d1-3B-w8a8: INT8 3.6GB
  • Available builds in LiquidAI/d1-3B-GGUF: Q4_K_M 1.6GB / Q8_0 2.7GB / BF16 5.0GB / F16 5.0GB
  • Available builds in LiquidAI/d1-3B-MLX-fp16: MLX 5.8GB
  • Available builds in LiquidAI/d1-3B-MLX-4bit: MLX 4bit 2.2GB
  • Available builds in LiquidAI/d1-3B-MLX-8bit: MLX 8bit 3.5GB

In addition, 22 converted build(s) from other uploaders exist on Hugging Face; this site lists only builds from the model’s publisher or established quantization maintainers.

This section is appended automatically by Local Model Watch when a converted build of this model appears after publication. Memory figures are estimated from the size of the distributed files. What the quantization names mean: glossary.

Related Articles

What to Read Next

Sources

Update History

  • 2026-11-17: Updated the hardware requirements table with the actual file sizes of LiquidAI/d1-3B-GGUF.
  • 2026-11-17: Changed the title to show what the article covers (VRAM requirements, file list, etc.).