GLM-5.3-Flash-Coder-RCO-152E-GGUF

GLM-5.3-Flash-Coder-RCO-152E-GGUF

Runs a pruned GLM-5.3-Flash coding and agent model on a home PC, as GGUF files in five quant sizes. It keeps 152 of 288 experts per layer, cutting a 321B model to about 170B.

Made by lapadatbogdan 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 repository holds GGUF files of GLM-5.3-Flash-Coder-RCO-152E, a pruned version of zai-org's GLM-5.3-Flash. It keeps 152 of the 288 experts in each layer, which shrinks the 321B mixture-of-experts model to about 170B with 17B active. The aim is coding and agent work on a desktop at home.

Who it suits: Developers with a strong desktop who want a local coding and agent model, and tinkerers who like to compare quantization levels.

Industry | Who uses it | How they use it | Example
Software | Developers | Run a large local coding model on a home PC | A coding assistant in a terminal
Agents | Agent builders | Back a tool-using agent with a local model | An agent that edits a project
Research | ML researchers | Study expert pruning and quantization trade-offs | A comparison of Q4_K_M and Q3_K_M
Hobby | Enthusiasts | Fit a large model onto a GPU plus system memory | A 24 GB card with 64 GB of RAM

If you have no large GPU or lots of memory, you can skip it, since even the smallest file is 47.4 GB.

Where to use it:

• llama.cpp and other programs that load GGUF files, on Windows, macOS or Linux
• Home PCs that pair a GPU with plenty of system RAM, or a Mac with large unified memory
• Multi-GPU machines, for the largest Q8_0 files

Use case | Who | What they get back
Coding help | Developers | Code written across many programming languages
Agentic coding | Agent builders | A model tuned on agentic sessions for tool use
Size comparison | Researchers | Five quant files from 47.4 GB to 180.4 GB

How to set it up:

• Python, to run the Hugging Face download tool
• Disk space for the quant you choose, from 47.4 GB to 180.4 GB
• Memory to match, such as 24 GB of VRAM plus 96 GB of RAM for Q4_K_M

1. Install the Hugging Face hub tools with pip.
2. Download the repository with the hf command, or pick one quant.
3. Check the SHA256 sums the card lists, since an earlier upload was truncated.
4. Load the first part of a split file in llama.cpp.

pip install -U huggingface_hub
hf download lapadatbogdan/GLM-5.3-Flash-Coder-RCO-152E-GGUF

How to use it: Load the quant that fits your memory in llama.cpp and use it for coding and agent tasks. The card says the Q3_K_M and Q4_K_M files include a merged fine-tune that removes reasoning loops, so the old logit-bias workaround is not needed.

Tool | What it does
Q8_0 | 180.4 GB in 4 parts, the reference quant
Q4_K_M | 102.6 GB in 3 parts, fits 24 GB VRAM plus 96 GB RAM
Q3_K_M | 75.1 GB in 2 parts, fits 24 GB VRAM plus 64 GB RAM
IQ2_XXS | 47.4 GB in 1 file, fits 24 GB VRAM plus 32 GB RAM
IQ2_XXS-downIQ2_S | 50.8 GB in 2 parts, fits 24 GB VRAM plus 32 GB RAM

Example: You load the Q4_K_M file and ask "Write a Python function that merges two sorted lists and add a short test". The model returns the function and a test, and the merged fine-tune keeps it from looping in its reasoning.

For newcomers and experts: Start with the Q4_K_M or Q3_K_M file if your memory allows it. Check the SHA256 sums before your first run, because a damaged download produces garbage output.

The pruning decided which experts stay by testing on agent sessions, code, math and everyday text, and a repair pass followed. The card measures drift against Q8_0 with KL divergence, where lower is closer. Q4_K_M picks the same top token as Q8_0 94.5% of the time, and Q3_K_M 92.5%.

Download a quant, checking SHA256
Load the first part in llama.cpp
Prompt for code or an agent task
Model answers without reasoning loops

Good to know:

• License: MIT
• Cost to run: none named, but the files are large and need a strong PC
• Data it reaches: downloads files from Hugging Face, then runs on your own machine
• Not for: small machines, since the smallest quant is 47.4 GB
• Limits: the card says an earlier upload was truncated and gave garbage output, so verify the sums

Title: Q3_K_M on 80 held-out coding runs, before and after the fine-tune
Looping runs before: 11 runs
Looping runs after: 0 runs
Solved before: 46 runs
Solved after: 51 runs

On Hugging Face, 13 November 2026

• Task: text generation
• Library: gguf
• Parameters: 170B
• Weights: about 84.8 GB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 34,478
• Likes: 36
• License: mit
• Based on: zai-org/GLM-5.3-Flash
• Tags: gguf, expert-pruning, code, agentic, llama.cpp, text-generation, conversational

Links

• Model: https://huggingface.co/lapadatbogdan/GLM-5.3-Flash-Coder-RCO-152E-GGUF
• Files: https://huggingface.co/lapadatbogdan/GLM-5.3-Flash-Coder-RCO-152E-GGUF/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download lapadatbogdan/GLM-5.3-Flash-Coder-RCO-152E-GGUF

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