GLM-5.3-Flash-oQ4e-mtp

GLM-5.3-Flash-oQ4e-mtp

A mixed-precision MLX copy of zai-org's GLM-5.3-Flash multimodal mixture-of-experts model, about 4.6 bits per weight, with the vision encoder and multi-token prediction head included.

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 GLM-5.3-Flash converted to MLX and compressed. The original is a mixture-of-experts model that handles text and images, with around 320B parameters overall and 18B in use for each token.

Three parts are bundled: the text model, an image encoder and a head that predicts several tokens at once.

The files were compressed with oQ, a mixed-precision method in oMLX, to an average of 4.641 bits per weight. The author says an earlier copy was re-quantized with a better low-bit method, so re-download if yours is old.

Who it suits: People with a large-memory machine who want to run a huge multimodal model locally through MLX, and people who study mixed-precision quantization.

Industry | Who uses it | How they use it | Example
Research | Quantization researchers | Study sensitivity-based bit allocation | Reading the card's bit distribution table
Software | Local-model developers | Run a large multimodal model on their own hardware | A private assistant that reads images and text
Hobbyists | MLX users | Try a frontier-scale open model locally | Testing it on a high-memory machine

You can skip it if your machine has far less than about 174 GiB of memory for the weights, or if you want a small model.

Where to use it: It is an MLX checkpoint, built for MLX tools and Apple's MLX framework on Apple silicon, and the card names oMLX as the quantization tool.

• MLX-based tools on Apple silicon
• oMLX, the tool that made the quantization
• A machine with room for roughly 174 GiB of weights

Use case | Who | What they get back
Local multimodal chat | Hobbyists | Replies to text and images from a huge model
Quantization study | Researchers | A documented mixed-precision layout
Speculative decoding experiments | 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 a lot of memory. The card gives only the download step and no run command.

• Python with pip
• An MLX-compatible runner
• About 174 GiB of free disk, and memory to match

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/GLM-5.3-Flash-oQ4e-mtp

How to use it: Load the folder in an MLX-based tool and chat as with any local model. The card includes no run command, so follow your tool's own steps.

Part | Storage format
Most weights | Affine 4-bit with group size 64
Sensitive layers | Affine 8-bit with group size 64
Some components | BF16
A few parameters | FP32

Example: You load the checkpoint in an MLX runner and ask it a question about a photo. The model answers from its language backbone and vision encoder, running only on your machine.

For newcomers and experts: If you are new, know that this is a huge model that needs a lot of memory. Start with a smaller checkpoint if your computer is modest.

If you know the field, oQ4e spends more bits on layers it measures as sensitive. Each group is then rounded using an importance matrix drawn from 128 samples of 512 tokens each, followed by a refit of the group's scale and bias.

Measure layer sensitivity
Assign bits per group
Round with an importance matrix
Refit scale and bias
Save the MLX checkpoint

Good to know: Nearly all weights are stored as 4-bit, and the language backbone dominates the files. The sizes come from the card.

Title: Storage by component, gigabytes
Language backbone: 180.966 GB
MTP head: 4.309 GB
Vision encoder: 1.127 GB

• License: MIT
• Cost to run: none named, but the weights need a large amount of memory
• Data it reaches: the files you download and the prompts and images you give it locally
• Not for: small machines, since the whole checkpoint is about 186 GB

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: mlx
• Parameters: 321B
• Weights: about 1,285 GB in U32
• Formats: Safetensors
• Downloads in the last 30 days: 26,944
• Likes: 7
• License: mit
• Based on: zai-org/GLM-5.3-Flash
• Tags: safetensors, glm5_next, quantized, image-text-to-text, conversational, 4-bit

Links

• Model: https://huggingface.co/Jundot/GLM-5.3-Flash-oQ4e-mtp
• Files: https://huggingface.co/Jundot/GLM-5.3-Flash-oQ4e-mtp/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Jundot/GLM-5.3-Flash-oQ4e-mtp

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