GLM-5.3-Flash Maya GGUF

GLM-5.3-Flash Maya GGUF

Five GGUF compressions of GLM-5.3-Flash, a 321B mixture-of-experts model, sized from 73.4 GB to 156.3 GB for Project Maya, which spreads the model across GPU, RAM and SSD.

Made by peasantsmith 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: GLM-5.3-Flash Maya GGUF is a set of five compressed versions of Z.ai's GLM-5.3-Flash. That model has 321B parameters and about 18B of them work on each token. The files are packed as GGUF for a tool called Project Maya.

Project Maya sorts the experts by how often they are used. The busiest sit on the GPU, a middle group sits in RAM and the least used wait on an SSD. This lets a PC run a model that is larger than its memory. More memory gives more speed, and the author asks you to test it on your own machine.

All five come from Z.ai's own FP8 release, not from a file that had already been compressed once. The folder set also holds the vision files.

Who it suits: It suits enthusiasts and developers who want to run a huge open model at home. The five sizes run from 73.4 GB to 156.3 GB.

Industry | Who uses it | How they use it | Example
Home labs | Hardware enthusiasts | Run a 321B model on a PC with spare RAM and SSD | Chatting with it on a desktop
Software | Developers | Use a large local model for code and tool calls | A coding helper that stays offline
Research | ML engineers | Compare compression levels against the original | Testing XS against L on one task
Gaming rigs | RTX 3090 or 4090 owners | Run the build tuned for 24 GB cards | Using the S24 folder

Skip it if your PC has little RAM, or you want a model that runs on a small GPU alone. The smallest build still needs about 64 GB of RAM plus a 24 GB card.

Where to use it:

• Project Maya, the runtime that splits the experts across GPU, RAM and SSD
• A PC with a 24 GB card, for the S24 build the card tunes for it
• Windows, macOS or Linux, with Python, for the file download alone

Use case | Who | What they get back
Large local model | Hobbyists | A 321B-class model running on a home PC
Tool calls and front-end code | Developers | The M build, which the card says is calibrated for them
Image input | Developers | The vision files that sit beside the five builds

How to set it up: You need a few things in place first.

• Python and pip, for the download command
• A PC with a large RAM pool and a GPU, plus a fast SSD for the overflow
• Project Maya, which the card says these files were made for
• Disk space for the folder you choose, from 73.4 GB up to 156.3 GB

1. Install the Hugging Face hub package with pip.
2. Download the repository with the hf command. It fetches all five folders, which add up to more than 500 GB, so fetch only the folder you need.
3. Point Project Maya at the folder you picked.

pip install -U huggingface_hub
hf download peasantsmith/GLM-5.3-Flash-Maya-GGUF

How to use it: You choose the version that fits your memory and load it in Project Maya. The visible part of the card has no run commands, so read the project's own notes for the exact steps.

Version | Size | Accuracy versus the full model | Built for
Maya-XS | 73.4 GB | 92.4 percent | 64 GB RAM and one 24 GB card
Maya-S | 96.5 GB | 97.9 percent | A smaller memory pool
Maya-S24 | 94.7 GB | 97.7 percent | 24 GB cards
Maya-M | 116 GB | 97.9 percent | A bigger memory pool
Maya-L | 156.3 GB | 99.2 percent | The biggest memory pool

Example: Your PC holds 64 GB of RAM and a single 24 GB graphics card. You download only the XS folder, load it in Project Maya and ask a coding question. The reply is slower than on a big server. The card says XS gives up some accuracy for its small size.

For newcomers and experts: Newcomers should try a smaller model first, since this setup takes real hardware and patience. If it fits your machine, pick S or S24 over XS when you can.

The accuracy figures compare each build with the original FP8 model on four zero-shot tests. The S24 build uses 4-bit for the attention and shared parts, which frees about 1.5 GB on the GPU for experts. The card says this gives up to 14 percent faster decoding on 24 GB cards.

Title: Accuracy versus the full FP8 model, from the card
Maya-L: 99.2 percent
Maya-M: 97.9 percent
Maya-S: 97.9 percent
Maya-S24: 97.7 percent
Maya-XS: 92.4 percent

Pick the build that fits your memory
Download that one folder
Load it in Project Maya
Experts spread over GPU, RAM and SSD

Good to know:

• License: MIT
• Cost to run: none named, but you need a lot of RAM, a GPU and an SSD
• Data it reaches: it downloads files from Hugging Face onto your machine
• Not for: machines with little memory, or work that needs the exact original quality, since even the closest build is below 100 percent

On Hugging Face, 13 November 2026

• Task: text generation
• Parameters: 4.1K
• Weights: about 1 MB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 1,276,300
• Likes: 109
• License: mit
• Based on: zai-org/GLM-5.3-Flash
• Quantized versions: 2
• Fine-tunes: 1
• Tags: gguf, imatrix, mixture-of-experts, project-maya, vision, text-generation

Links

• Model: https://huggingface.co/peasantsmith/GLM-5.3-Flash-Maya-GGUF
• Files: https://huggingface.co/peasantsmith/GLM-5.3-Flash-Maya-GGUF/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download peasantsmith/GLM-5.3-Flash-Maya-GGUF

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