LFM2-VL-3B

Highest-capacity 3B vision-language model with enhanced visual reasoning

← Back to Vision Models

LFM2-VL-3B is Liquid AI’s highest-capacity multimodal model, delivering enhanced visual reasoning and detailed image understanding. Ideal for complex vision tasks requiring deeper comprehension.

HFGGUFMLXONNX

Specifications

Property	Value
Parameters	3B
Context Length	32K tokens
Architecture	LFM2-VL (Dense)

Advanced Reasoning

Complex visual logic and analysis

Document Understanding

Detailed document and chart parsing

Multi-Image

Compare and reason across images

Quick Start

Transformers

Install:

uv pip install "transformers>=5.0.0" pillow torch

Download & Run:

from transformers import AutoProcessor, AutoModelForImageTextToText
from transformers.image_utils import load_image

model_id = "LiquidAI/LFM2-VL-3B"
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    device_map="auto",
    dtype="bfloat16",
)
# IMPORTANT: tie lm_head to input embeddings (transformers v5 bug)
model.lm_head.weight = model.get_input_embeddings().weight

processor = AutoProcessor.from_pretrained(model_id)

url = "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
image = load_image(url)

conversation = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": image},
            {"type": "text", "text": "What is in this image?"},
        ],
    },
]

inputs = processor.apply_chat_template(
    conversation,
    add_generation_prompt=True,
    return_tensors="pt",
    return_dict=True,
    tokenize=True,
).to(model.device)

outputs = model.generate(**inputs, do_sample=True, temperature=0.1, min_p=0.15, repetition_penalty=1.05, max_new_tokens=256)
response = processor.batch_decode(outputs, skip_special_tokens=True)[0]
print(response)

vLLM

vLLM support for LFM Vision Models requires a specific version. Install from the custom source below.

Install:

VLLM_PRECOMPILED_WHEEL_COMMIT=72506c98349d6bcd32b4e33eec7b5513453c1502 \
  VLLM_USE_PRECOMPILED=1 \
  uv pip install git+https://github.com/vllm-project/vllm.git

uv pip install "transformers>=5.0.0" pillow

Run:

from vllm import LLM, SamplingParams

IMAGE_URL = "http://images.cocodataset.org/val2017/000000039769.jpg"

llm = LLM(
    model="LiquidAI/LFM2-VL-3B",
    max_model_len=1024,
)

sampling_params = SamplingParams(
    temperature=0.1,
    min_p=0.15,
    repetition_penalty=1.05,
    max_tokens=256,
)

messages = [{
    "role": "user",
    "content": [
        {"type": "image_url", "image_url": {"url": IMAGE_URL}},
        {"type": "text", "text": "Describe what you see in this image."},
    ],
}]

outputs = llm.chat(messages, sampling_params)
print(outputs[0].outputs[0].text)

llama.cpp

llama.cpp enables efficient CPU inference for vision models.

Install:

brew install llama.cpp

Or download pre-built binaries from llama.cpp releases.

Run:

llama-cli \
    -hf LiquidAI/LFM2-VL-3B-GGUF:Q4_0 \
    --image test_image.jpg \
    -p "What's in this image?" \
    -n 128 \
    --temp 0.1 --min-p 0.15 --repeat-penalty 1.05

The -hf flag downloads the model directly from Hugging Face. Use --image-max-tokens to control image token budget.

For server deployment and advanced usage, see the llama.cpp guide.