Qwen3.6-35B-A3B-8bit (MLX)

Qwen3.6-35B-A3B-8bit (MLX)

8-bit MLX conversion of Qwen3.6-35B-A3B that reads text and images and answers in text, run with the mlx-vlm tool. For developers who want to run a large vision-language model locally.

Made by mlx-community on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This is an 8-bit copy of Qwen3.6-35B-A3B in MLX format, which reads both text and images and answers in text. It was converted with mlx-vlm version 0.4.4. The model has 35.1 billion parameters. For details about the model itself, the card sends you to the original Qwen3.6-35B-A3B card.

Who it suits: It suits developers and tinkerers who want to run a large vision-language model through the MLX toolchain on their own computer.

Industry | Who uses it | How they use it | Example
Software | App developers | Add image description to a local app | Describing screenshots from a script
Research | Model testers | Try an 8-bit conversion of a large model | Comparing replies with another quantization
Accessibility | Tool makers | Generate text descriptions of pictures | Writing alt text for a photo

You can skip it if you do not use MLX, or if you want the original full-precision model.

Where to use it: It runs with the mlx-vlm command line tool and Python, on computers that support MLX, which is usually Apple hardware.

• The mlx-vlm tool in a terminal
• Python scripts that use mlx-vlm
• Windows, macOS or Linux, for downloading the files

Use case | Who | What they get back
Image description | Developers | A text description of a picture you pass in
Visual question answering | Researchers | Written answers about an image
Local chat with images | Hobbyists | Replies produced entirely on your computer

How to set it up: You need Python and the mlx-vlm package. Hugging Face lists the files at about 140 GB, so check free disk space first. The model downloads by its name the first time you run the command.

• Python
• The mlx-vlm package
• Plenty of free disk space

1. Install mlx-vlm with pip.
2. Run the generate command with the model name and an image path.

pip install -U mlx-vlm

How to use it: Run the generate module with the model name, a token limit, a temperature, a prompt and the path to an image. Setting the temperature to 0.0 gives the same reply each time.

python -m mlx_vlm.generate --model mlx-community/Qwen3.6-35B-A3B-8bit --max-tokens 100 --temperature 0.0 --prompt "Describe this image." --image <path_to_image>

Tool | What it does
mlx_vlm.generate | Runs the model on a prompt and an image
--max-tokens | Limits the length of the reply
--temperature | Controls how random the reply is, 0.0 for steady output
--image | Gives the path to the picture to describe

Example: You run the command with the prompt "Describe this image." and the path to a photo. The model prints a short text description of what the picture shows, up to 100 tokens long.

For newcomers and experts: If you are new, install mlx-vlm, pick a photo and run the one command above. Read the printed description to see what the model noticed.

If you know the field, this is a quantized conversion for MLX of a mixture-of-experts model from the Qwen3.6 series. The card adds no benchmark data of its own, so refer to the original model's card for capabilities and limits.

Photo and a text prompt
mlx-vlm loads the 8-bit model
Model reads the image and prompt
Text reply printed in the terminal

Good to know: The card is short and relies on the original Qwen3.6-35B-A3B card for everything else.

• License: Apache 2.0
• Cost to run: no API key, account or paid service is named, but the files are large
• Data it reaches: the image and prompt you give it on your machine, plus files downloaded from Hugging Face
• Not for: setups without MLX support

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: mlx
• Parameters: 35.1B
• Weights: about 140 GB in U32
• Formats: Safetensors
• Downloads in the last 30 days: 5,687
• Likes: 37
• License: apache-2.0
• Based on: Qwen/Qwen3.6-35B-A3B
• Quantized versions: 23
• Tags: safetensors, qwen3_5_moe, image-text-to-text, conversational, 8-bit

Links

• Model: https://huggingface.co/mlx-community/Qwen3.6-35B-A3B-8bit
• Files: https://huggingface.co/mlx-community/Qwen3.6-35B-A3B-8bit/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download mlx-community/Qwen3.6-35B-A3B-8bit

Then follow the usage in its model card. Nothing in this kit runs by itself.