Jan-v2-VL-high 8-bit MLX

This is an 8-bit quantized MLX conversion of janhq/Jan-v2-VL-high.

Model Description

Jan-v2-VL is an 8-billion parameter vision-language model designed for long-horizon, multi-step tasks in real software environments. This "high" variant is optimized for deeper reasoning and complex task execution, providing the highest quality outputs 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 8-bit quantization using MLX-VLM by Prince Canuma.

Conversion command:

mlx_vlm.convert --hf-path janhq/Jan-v2-VL-high --quantize --q-bits 8 --mlx-path Jan-v2-VL-high-8bit-mlx

Usage

Installation

pip install mlx-vlm

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-high-8bit-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-high-8bit-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-high

Acknowledgments

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