Try LFM • Docs • LEAP • Discord

LFM2.5‑VL-450M

LFM2.5‑VL-450M is Liquid AI's refreshed version of the first vision-language model, LFM2-VL-450M, built on an updated backbone LFM2.5-350M and tuned for stronger real-world performance. Find more about LFM2.5 family of models in our blog post.

Enhanced instruction following on vision and language tasks.
Improved multilingual vision understanding in Arabic, Chinese, French, German, Japanese, Korean, Portuguese and Spanish.
Bounding box prediction and object detection for grounded visual understanding.
Function calling support for text-only input.

🎥⚡️ You can try LFM2.5-VL-450M running locally in your browser with our real-time video stream captioning WebGPU demo 🎥⚡️

Alternatively, try the API model on the Playground.

📄 Model details

LFM2.5-VL-450M is a general-purpose vision-language model with the following features:

LM Backbone: LFM2.5-350M
Vision encoder: SigLIP2 NaFlex shape‑optimized 86M
Context length: 32,768 tokens
Vocabulary size: 65,536
Languages: English, Arabic, Chinese, French, German, Japanese, Korean, Portuguese, and Spanish
Native resolution processing: handles images up to 512*512 pixels without upscaling and preserves non-standard aspect ratios without distortion
Tiling strategy: splits large images into non-overlapping 512×512 patches and includes thumbnail encoding for global context
Inference-time flexibility: user-tunable maximum image tokens and tile count for speed/quality tradeoff without retraining
Generation parameters:
- text: temperature=0.1, min_p=0.15, repetition_penalty=1.05
- vision: min_image_tokens=32 max_image_tokens=256, do_image_splitting=True

Model	Description
LFM2.5-VL-450M	Original model checkpoint in native format. Best for fine-tuning or inference with Transformers and vLLM.
LFM2.5-VL-450M-GGUF	Quantized format for llama.cpp and compatible tools. Optimized for CPU inference and local deployment with reduced memory usage.
LFM2.5-VL-450M-ONNX	ONNX Runtime format for cross-platform deployment. Enables hardware-accelerated inference across diverse environments (cloud, edge, mobile).
LFM2.5-VL-450M-MLX-8bit	MLX format for Apple Silicon. Optimized for fast on-device inference on Mac with mlx-vlm. Also available in 4bit, 5bit, 6bit, and bf16.

We recommend using it for general vision-language workloads, captioning and object detection. It’s not well-suited for knowledge-intensive tasks or fine-grained OCR.

Chat Template

LFM2.5-VL uses a ChatML-like format. See the Chat Template documentation for details.

<|startoftext|><|im_start|>system
You are a helpful multimodal assistant by Liquid AI.<|im_end|>
<|im_start|>user
<image>Describe this image.<|im_end|>
<|im_start|>assistant
This image shows a Caenorhabditis elegans (C. elegans) nematode.<|im_end|>

You can use processor.apply_chat_template() to format your messages automatically.

🏃 Inference

You can run LFM2.5-VL-450M with Hugging Face transformers v5.1 or newer:

pip install transformers pillow

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

# Load model and processor
model_id = "LiquidAI/LFM2.5-VL-450M"
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    device_map="auto",
    dtype="bfloat16"
)
processor = AutoProcessor.from_pretrained(model_id)

# Load image and create conversation
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?"},
        ],
    },
]

# Generate Answer
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, max_new_tokens=64)
processor.batch_decode(outputs, skip_special_tokens=True)[0]

# This image captures the iconic Statue of Liberty standing majestically on Liberty Island in New York City. The statue, a symbol of freedom and democracy, is prominently featured in the foreground, its greenish-gray hue contrasting beautifully with the surrounding water.

Visual grounding

LFM2.5-VL-450M supports bounding box prediction:

url = "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
image = load_image(url)
query = "status"
prompt = f'Detect all instances of: {query}. Response must be a JSON array: [{"label": ..., "bbox": [x1, y1, x2, y2]}, ...]. Coordinates are normalized to [0,1].'

conversation = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": image},
            {"type": "text", "text": prompt},
        ],
    },
]

# Generate Answer
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, max_new_tokens=64)
processor.batch_decode(outputs, skip_special_tokens=True)[0]

# [{"label": "statue", "bbox": [0.3, 0.25, 0.4, 0.65]}]

Tool Use

LFM2.5 supports function calling for text only input by applying the chat template with the tokenizer. See the Tool Use documentation for the full guide.

tools = [{
    "name": "get_weather",
    "description": "Get current weather for a location",
    "parameters": {
        "type": "object",
        "properties": {"location": {"type": "string"}},
        "required": ["location"]
    }
}]

messages = [{"role": "user", "content": "What's the weather in Paris?"}]

# Apply chat template with tools
inputs = processor.tokenizer.apply_chat_template(
    messages,
    tools=tools,
    add_generation_prompt=True,
    return_tensors="pt",
    return_dict=True,
)
input_ids = inputs["input_ids"].to(model.device)
outputs = model.generate(input_ids, max_new_tokens=256)
response = processor.tokenizer.decode(outputs[0, input_ids.shape[1]:], skip_special_tokens=False)

# <|tool_call_start|>[get_weather(location="Paris")]<|tool_call_end|>I am retrieving the current weather for Paris.<|im_end|>

Name	Description	Docs
Transformers	Simple inference with direct access to model internals.	Link
vLLM	High-throughput production deployments with GPU.	Link
SGLang	High-throughput production deployments with GPU.	Link
llama.cpp	Cross-platform inference with CPU offloading.	Link

🔧 Fine-tuning

We recommend fine-tuning LFM2.5-VL-450M model on your use cases to maximize performance.

Notebook	Description	Link
SFT (Unsloth)	Supervised Fine-Tuning with LoRA using Unsloth.
SFT (TRL)	Supervised Fine-Tuning with LoRA using TRL.

📊 Performance

LFM2.5-VL-450M improves over LFM2-VL-450M across both vision and language benchmarks, while also adding two new capabilities: bounding box prediction on RefCOCO-M and function calling support measured by BFCLv4.

Vision benchmarks

Model	MMStar	RealWorldQA	MMBench (dev en)	MMMU (val)	POPE	MMVet	BLINK	InfoVQA (val)	OCRBench	MM-IFEval	MMMB	CountBench	RefCOCO-M
LFM2.5-VL-450M	43.00	58.43	60.91	32.67	86.93	41.10	43.92	43.02	684	45.00	68.09	73.31	81.28
LFM2-VL-450M	40.87	52.03	56.27	34.44	83.79	33.85	42.61	44.56	657	33.09	54.29	47.64	-
SmolVLM2-500M	38.20	49.90	52.32	34.10	82.67	29.90	40.70	24.64	609	11.27	46.79	61.81	-

All vision benchmark scores are obtained using VLMEvalKit. Multilingual scores are based on the average of benchmarks translated by GPT-4.1-mini from English to Arabic, Chinese, French, German, Japanese, Korean, Portuguese, and Spanish.

Language benchmarks

Model	GPQA	MMLU Pro	IFEval	Multi-IF	BFCLv4
LFM2.5-VL-450M	25.66	19.32	61.16	34.63	21.08
LFM2-VL-450M	23.13	17.22	51.75	26.21	-
SmolVLM2-500M	23.84	13.57	30.14	6.82	-

📬 Contact

Got questions or want to connect? Join our Discord community
If you are interested in custom solutions with edge deployment, please contact our sales team.

Citation

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

LiquidAI/LFM2.5-VL-450M