> ## Documentation Index
> Fetch the complete documentation index at: https://docs.liquid.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# LFM2-VL-1.6B

> 1.6B vision-language model (deprecated - use LFM2.5-VL-1.6B instead)

[← Back to Vision Models](/lfm/models/vision-models.md)

> [!WARNING]
>
> This model is deprecated. Use [LFM2.5-VL-1.6B](/lfm/models/lfm25-vl-1.6b) for improved performance.

LFM2-VL-1.6B was the original 1.6B vision-language model. It has been superseded by LFM2.5-VL-1.6B, which is built on the newer LFM2.5-1.2B-Base language backbone and improves instruction following, multilingual vision understanding, and OCR.

[HF](https://huggingface.co/LiquidAI/LFM2-VL-1.6B)
[GGUF](https://huggingface.co/LiquidAI/LFM2-VL-1.6B-GGUF)
[MLX](https://huggingface.co/mlx-community/LFM2-VL-1.6B-8bit)
[ONNX](https://huggingface.co/onnx-community/LFM2-VL-1.6B-ONNX)

## Specifications

| Property | Value |
| - | - |
| Parameters | 1.6B |
| Context Length | 32K tokens |
| Architecture | LFM2-VL (Dense) |
| Languages | English |
| License | [LFM Open License v1.0](/lfm/help/model-license.md) |

## Quick Start

**Transformers**

**Install:**

pip install "transformers>=5.2.0" pillow torch torchvision accelerate

**Download & Run:**

from transformers import AutoProcessor, AutoModelForImageTextToText
from transformers.image\_utils import load\_image

model\_id = "LiquidAI/LFM2-VL-1.6B"
model = AutoModelForImageTextToText.from\_pretrained(
model\_id,
device\_map="auto",
dtype="bfloat16",
)

processor = AutoProcessor.from\_pretrained(model\_id)

url = "[https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg](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)
input\_length = inputs\["input\_ids"].shape\[1]
response = processor.batch\_decode(outputs\[:, input\_length:], skip\_special\_tokens=True)\[0]
print(response)

**vLLM**

**Install:**

uv pip install vllm==0.19.0

uv pip install transformers==5.5.0 pillow

**Run:**

from vllm import LLM, SamplingParams

IMAGE\_URL = "[http://images.cocodataset.org/val2017/000000039769.jpg](http://images.cocodataset.org/val2017/000000039769.jpg)"

llm = LLM(
model="LiquidAI/LFM2-VL-1.6B",
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)

**SGLang**

**Install:**

uv pip install "sglang>=0.5.10"

**Launch server:**

sglang serve \\
\--model-path LiquidAI/LFM2-VL-1.6B \\
\--host 0.0.0.0 \\
\--port 30000

**Query (OpenAI-compatible):**

from openai import OpenAI

client = OpenAI(base\_url="[http://localhost:30000/v1](http://localhost:30000/v1)", api\_key="None")

response = client.chat.completions.create(
model="LiquidAI/LFM2-VL-1.6B",
messages=\[\{
"role": "user",
"content": \[
\{"type": "image\_url", "image\_url": \{"url": "[http://images.cocodataset.org/val2017/000000039769.jpg](http://images.cocodataset.org/val2017/000000039769.jpg)"}},
\{"type": "text", "text": "Describe what you see in this image."},
],
}],
temperature=0.1, extra\_body=\{"min\_p": 0.15, "repetition\_penalty": 1.05},
max\_tokens=256,
)

print(response.choices\[0].message.content)

**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](https://github.com/ggml-org/llama.cpp/releases).

**Run:**

llama-cli \\
-hf LiquidAI/LFM2-VL-1.6B-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](/deployment/on-device/llama-cpp.md#vision-models).


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.