Jan-v2-VL is an 8-billion parameter vision-language model designed for long-horizon, multi-step tasks in real software environments. This "med" variant provides a balanced trade-off between inference speed and reasoning capability, offering strong performance for agentic automation and UI control tasks.

Key Features:

Vision-language understanding for browser and desktop applications
Screenshot grounding and tool call capabilities
Stable multi-step execution with minimal performance drift
Error recovery and intermediate state maintenance

Quantization

This model was converted to MLX format with 4-bit quantization using MLX-VLM by Prince Canuma.

Conversion command:

mlx_vlm.convert --hf-path janhq/Jan-v2-VL-med --quantize --q-bits 4 --mlx-path Jan-v2-VL-med-4bit-mlx

Usage

Installation

pip install mlx-vlm

Python

python

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

# Load the model
model_path = "mlx-community/Jan-v2-VL-med-4bit-mlx"
model, processor = load(model_path)
config = load_config(model_path)

# Prepare input
image = ["path/to/image.jpg"]
prompt = "Describe this image."

# Apply chat template
formatted_prompt = apply_chat_template(
    processor, config, prompt, num_images=len(image)
)

# Generate output
output = generate(model, processor, formatted_prompt, image, verbose=False)
print(output)

Command Line

mlx_vlm.generate --model mlx-community/Jan-v2-VL-med-4bit-mlx --max-tokens 100 --prompt "Describe this image" --image path/to/image.jpg

Intended Use

This model is designed for:

Agentic automation and UI control
Stepwise operation in browsers and desktop applications
Screenshot grounding and tool calls
Long-horizon multi-step task execution

License

This model is released under the Apache 2.0 license.

Original Model

For more information, please refer to the original model: janhq/Jan-v2-VL-med

Acknowledgments

Original model by Jan
MLX framework by Apple
MLX conversion framework by Prince Canuma
Model conversion by Incept5