LFM3-VL-2B

← Back to Vision Models LFM3-VL-2B is a vision-language model in the LFM3 family designed for on-device deployment. It processes text, images, and videos at their native resolution, and pairs the LFM3 language backbone with a new in-house Liquid hybrid convolution/attention vision encoder. It is the first LFM vision model to take video as input, and it reasons step by step before answering.

HF GGUF MLX ONNX

Specifications

Property	Value
Parameters	2B (1.73B language model + 164M vision encoder + 25M projector)
Architecture	LFM3-VL — LFM3 backbone + Liquid hybrid vision encoder, 14 layers (8 gated depthwise 2D convolution + 6 bidirectional attention)
Context Length	128K tokens
Languages	English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Spanish, Vietnamese, Thai, Indonesian, Hindi, Russian, Polish
Vocabulary Size	128K
Generation Parameters	`temperature=0.1`, `top_k=50`, `repetition_penalty=1.1`
License	LFM Open License v1.0

Video Understanding

Questions about recorded lectures, phone walkthroughs, and camera clips

Thinking

Step-by-step reasoning for math, charts, and multi-step visual questions

Spatial & Embodied Reasoning

Object localization and spatial relationships for robotics applications

Quick Start

Transformers
vLLM
SGLang
llama.cpp

Install:

pip install "transformers>=5.21.0" torch torchvision pillow

Download & Run:

import torch
from transformers import AutoModelForImageTextToText, AutoProcessor

model_id = "LiquidAI/LFM3-VL-2B"

model = AutoModelForImageTextToText.from_pretrained(model_id, dtype=torch.bfloat16, device_map="auto", trust_remote_code=True)
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)

messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "http://images.cocodataset.org/val2017/000000039769.jpg"},
            {"type": "text", "text": "Describe this image."},
        ],
    }
]

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

with torch.inference_mode():
    output_ids = model.generate(
        **inputs,
        do_sample=True,
        temperature=0.1,
        top_k=50,
        repetition_penalty=1.1,
        max_new_tokens=16384,
    )

generated_ids = output_ids[:, inputs["input_ids"].shape[1] :]
print(processor.batch_decode(generated_ids, skip_special_tokens=True)[0])

Requires a vLLM build with LFM3-VL support.Launch server:

vllm serve LiquidAI/LFM3-VL-2B \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice --tool-call-parser qwen3_coder \
  --chat-template-content-format openai \
  --structured-outputs-config '{"backend": "xgrammar", "disable_any_whitespace": true}' \
  --media-io-kwargs '{"video": {"video_backend": "qwen3_vl"}}'

Query (OpenAI-compatible):

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")

response = client.chat.completions.create(
    model="LiquidAI/LFM3-VL-2B",
    messages=[
        {
            "role": "user",
            "content": [
                {"type": "image_url", "image_url": {"url": "http://images.cocodataset.org/val2017/000000039769.jpg"}},
                {"type": "text", "text": "Describe this image."},
            ],
        }
    ],
    temperature=0.1,
    extra_body={"top_k": 50, "repetition_penalty": 1.1},
)
print(response.choices[0].message.content)

Requires an SGLang build with LFM3-VL support.Launch server:

sglang serve --model-path LiquidAI/LFM3-VL-2B \
  --reasoning-parser qwen3 \
  --tool-call-parser qwen3_coder \
  --constrained-json-disable-any-whitespace \
  --host 0.0.0.0 --port 30000

Query (OpenAI-compatible):

from openai import OpenAI

client = OpenAI(base_url="http://localhost:30000/v1", api_key="None")

response = client.chat.completions.create(
    model="LiquidAI/LFM3-VL-2B",
    messages=[
        {
            "role": "user",
            "content": [
                {"type": "image_url", "image_url": {"url": "http://images.cocodataset.org/val2017/000000039769.jpg"}},
                {"type": "text", "text": "Describe this image."},
            ],
        }
    ],
    temperature=0.1,
    extra_body={"top_k": 50, "repetition_penalty": 1.1},
)
print(response.choices[0].message.content)

Requires a llama.cpp build with LFM3-VL support.Install:

brew install llama.cpp

Start the server:

llama-server \
  -hf LiquidAI/LFM3-VL-2B-GGUF:Q4_0 \
  --temp 0.1 --top-p 1.0 --top-k 50 --min-p 0.0 --repeat-penalty 1.1 \
  --no-reasoning-preserve --port 8080

With -hf, llama.cpp downloads the model and the matching vision projector (mmproj). The server is OpenAI-compatible, so you can query it with the vLLM example above by changing base_url to http://localhost:8080/v1. To chat in the terminal, use llama-cli with the same flags, then type /image <path> to add an image to the chat.For advanced usage, see the llama.cpp guide.

Reasoning

LFM3-VL-2B thinks in a <think> block before it answers. Reasoning can be long, so allow up to 16,384 new tokens. See Chat Template and Tool Use for the message format.

Video Input

Images and videos are processed at their native resolution, and every 2×2 block of patches is merged into one language token. Videos are sampled at 2 fps, consecutive frames are paired, and each frame is prefixed with its timestamp. See Vision Capabilities for a video example.

Models

Fine-tuning

Edge Inference

GPU Inference

Cloud Inference

Guides

Help

Specifications

Video Understanding

Thinking

Spatial & Embodied Reasoning

Quick Start

Reasoning

Video Input

​Specifications

Video Understanding

Thinking

Spatial & Embodied Reasoning

​Quick Start

​Reasoning

​Video Input

Specifications

Quick Start

Reasoning

Video Input