LFM2.5-VL-3B is a vision-language model built on the LFM2.5-2.6B backbone with a SigLIP2 NaFlex vision encoder (400M). It supports OCR, document comprehension, multilingual vision understanding, bounding box prediction, and function calling.

Quickstart

uv run --with mlx-vlm mlx_vlm.generate --model LiquidAI/LFM2.5-VL-3B-MLX-8bit --max-tokens 100 --temperature 0.2 --image https://placecats.com/neo/300/200 --prompt "how many animals are in the picture?"

python

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

model, processor = load("LiquidAI/LFM2.5-VL-3B-MLX-8bit")

image = load_image("https://placecats.com/neo/300/200")

messages = [
    {
        "role": "user",
        "content": [
            {"type": "image"},
            {"type": "text", "text": "What do you see in this image?"},
        ],
    }
]
prompt = apply_chat_template(
    processor,
    model.config,
    messages,
    add_generation_prompt=True,
    num_images=1,
)

result = generate(
    model,
    processor,
    prompt,
    [image],
    temp=0.2,
    top_k=50,
    repetition_penalty=1.0,
    verbose=True,
)
print(result.text)