LFM2.5-VL-450M-Extract-MLX-4bit is the 4-bit quantized Apple Silicon MLX port of Liquid AI's LFM2.5-VL-450M-Extract. It extracts user-defined fields from images and returns them as strict JSON, optimized for ultra-low latency and minimal memory consumption on Apple Silicon (M1/M2/M3/M4) devices via MLX and mlx-vlm.

⚙️ How it works

You specify what to extract as a YAML field list in the system prompt, and the model returns a JSON object with those fields.

System prompt:

yaml

wood_color: The overall coloration of the wood surface
wood_texture: The tactile quality of the wood surface 
wood_pattern: The pattern types visible on the wood surface

Output:

json

{
  "wood_color": "light to medium brown",
  "wood_texture": "smooth with visible grain",
  "wood_pattern": "parallel, irregular, wavy"
}

Our model also supports enum constrained definitions:

yaml

wood_color: The overall coloration of the wood surface, such as blue, red, or light tan
wood_texture: The tactile quality of the wood surface, select from smooth, rough, or grainy
wood_pattern: The pattern types visible on the wood surface, e.g., straight, wavy, or curly

📄 Model Details

Property	Detail
Base Model	LiquidAI/LFM2.5-VL-450M-Extract
Parameters (LM only)	350M
Vision Encoder	SigLIP2 (~100M, SigLIP-2 paper)
Quantization	4-bit affine (`group_size: 64`, `mode: affine`)
Weights Size	~388 MB (`model.safetensors`)
Context Window	128,000 tokens
Runtime / Library	Apple Silicon MLX (`mlx_vlm`)
License	LFM Open License v1.0

📊 Conversion Specs & Resource Usage

Conversion Parameters: python -m mlx_vlm.convert --hf-path LiquidAI/LFM2.5-VL-450M-Extract -q --q-bits 4 --q-group-size 64
Memory Footprint: ~400 MB active unified memory during inference.

Model Variant	Precision	Group Size	Weight Size	Est. Unified RAM
LFM2.5-VL-450M-Extract-MLX-4bit	4-bit (`int4`)	64	~388 MB	~400 MB
LFM2.5-VL-450M-Extract-MLX-8bit	8-bit (`int8`)	64	~565 MB	~600 MB
LFM2.5-VL-450M-Extract-MLX-bf16	bfloat16	N/A	~897 MB	~900 MB

🏃 How to run with MLX

1. Install dependencies

pip install mlx-vlm pillow

2. Run inference in Python

python

from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template
from mlx_vlm.utils import load_image

model_id = "LeoFortunato/LFM2.5-VL-450M-Extract-MLX-4bit"
model, processor = load(model_id)

image = load_image("https://huggingface.co/LiquidAI/LFM2.5-VL-450M-Extract/resolve/main/sample_image.png")

fields_yaml = """wood_color: The overall coloration of the wood surface
wood_texture: The tactile quality of the wood surface
wood_pattern: The pattern types visible on the wood surface"""

system_prompt = f"""Extract the following from the image:

{fields_yaml}

Respond with only a JSON object. Do not include any text outside the JSON."""

formatted_prompt = apply_chat_template(
    processor,
    processor.config,
    system_prompt,
    num_images=1
)

output = generate(
    model,
    processor,
    formatted_prompt,
    image=[image],
    max_tokens=512,
    temperature=0.0,
    verbose=False
)
print(output)

📬 Citation & Contact

Liquid AI Discord: Join Discord
MLX Framework: MLX GitHub

bibtex

@article{liquidai2025lfm2,
 title={LFM2 Technical Report},
 author={Liquid AI},
 journal={arXiv preprint arXiv:2511.23404},
 year={2025}
}