Qwen3.6-35B-A3B-oQ6e-mtp

Qwen3.6-35B-A3B-oQ6e-mtp

A mixed-precision MLX copy of Qwen3.6-35B-A3B at about 6.8 bits per weight, with the vision encoder and multi-token prediction head included, for running locally on Apple silicon.

Made by Jundot 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 Qwen3.6-35B-A3B converted to MLX and compressed with oQ6e, a mixed-precision method from oMLX. The original is a mixture-of-experts model for text and images, with 35B parameters overall and 3B in use per token.

Its layers alternate between Gated DeltaNet and Gated Attention designs. Three parts are bundled: the language model, the vision encoder and a head for predicting several tokens at once. The whole checkpoint averages 6.762 bits per weight.

Who it suits: People who run models locally with MLX and want a near-full-quality copy of a mid-sized multimodal model that fits in about 28 GiB. Quantization researchers can also read its bit table.

Industry | Who uses it | How they use it | Example
Software | Local-model developers | Run a multimodal model on their own machine | An offline assistant that reads screenshots
Research | Quantization researchers | Study mixed 6-bit and 8-bit allocation | Reading the card's bit distribution table
Hobbyists | MLX users | Chat with a mid-sized model locally | A private assistant on a Mac

You can skip it if you want a much smaller file, since this copy keeps high precision, or if you do not use MLX.

Where to use it: It is an MLX checkpoint for MLX-based tools on Apple silicon. The card names oMLX as the tool that made the quantization.

• MLX-based tools on Apple silicon
• oMLX, the quantization tool behind it
• Machines with room for roughly 28 GiB of weights

Use case | Who | What they get back
Local chat with images | Hobbyists | Replies to text and picture prompts from a local model
Higher-fidelity quantization | Researchers | A 6-bit copy to compare with lower-bit ones
Multi-token prediction tests | Developers | A checkpoint that includes the MTP head

How to set it up: You need Python for the download tool, an MLX-based runner and about 28 GiB of space. The card gives only the download step and no run command.

• Python with pip
• An MLX-compatible runner
• About 28.3 GiB of disk space

1. Install the Hugging Face hub tools with pip.
2. Download the repository files with the hf command.
3. Open the checkpoint in your MLX runner.

pip install -U huggingface_hub
hf download Jundot/Qwen3.6-35B-A3B-oQ6e-mtp

How to use it: Load the folder in an MLX-based tool and chat with it. The card gives no run command, so follow the steps of the tool you use. The card says to re-download if you hold an older copy of this model.

Part | Storage format
Most weights | Affine 6-bit with group size 64
Some layers | Affine 8-bit with group size 64 or 128
Other components | BF16

Example: You open the checkpoint in an MLX runner and ask for a summary of a chart image. The model replies from its language backbone and vision encoder, entirely on your machine.

For newcomers and experts: If you are new, this is a ready-to-load model folder. Download it with the two commands above, then open it in an MLX tool and start with a short text question.

If you know the field, oQ6e spends bits according to measured layer sensitivity. It rounds each group with an importance matrix drawn from 128 samples of 512 tokens, then refits scale and bias. The language backbone averages 6.645 bits per weight.

Measure layer sensitivity
Assign 6-bit or 8-bit per group
Round with an importance matrix
Refit scale and bias
Save the MLX checkpoint

Good to know: Over 90 percent of weights are stored as 6-bit. The card gives the sizes below for the total of about 30.388 GB.

Title: Storage by component, gigabytes
Language backbone: 28.791 GB
Vision encoder: 0.893 GB
MTP head: 0.704 GB

• License: Apache 2.0
• Cost to run: none named
• Data it reaches: the files you download, plus the prompts and images you give it locally
• Not for: tools outside MLX, since this is an MLX checkpoint

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: mlx
• Parameters: 36B
• Weights: about 144 GB in U32
• Formats: Safetensors
• Downloads in the last 30 days: 1,052
• Likes: 4
• License: apache-2.0
• Based on: Qwen/Qwen3.6-35B-A3B
• Tags: safetensors, qwen3_5_moe, quantized, image-text-to-text, conversational, 6-bit

Links

• Model: https://huggingface.co/Jundot/Qwen3.6-35B-A3B-oQ6e-mtp
• Files: https://huggingface.co/Jundot/Qwen3.6-35B-A3B-oQ6e-mtp/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Jundot/Qwen3.6-35B-A3B-oQ6e-mtp

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