GitHub radar
A GLM-5.3-class AI model that fits on one graphics card
GLM-5.3-SLIM-GGUF is a trimmed version of the large GLM-5.3 language model: a quarter of its internal "expert" blocks were removed without retraining, so it now fits on a single powerful GPU or Mac Studio instead of a data center.
It's a large language model (AI like ChatGPT) with part of its internal building blocks removed, so it needs less memory and weaker hardware to run.
When it helps
Useful if you own an NVIDIA B200 GPU or a Mac Studio with 192 GB or more of memory and want to run a powerful language model locally; it won't run on a regular laptop or home PC.
Pros
- 24.6% smaller than the original model with barely any quality loss
- Fits on one GPU instead of a whole server cluster
- Scores 92.7% on the HumanEval coding benchmark
- Step-by-step setup instructions are on the model page
Cons
- Needs an NVIDIA B200 GPU or a 192 GB+ Mac Studio — ordinary hardware won't do
- Quality was measured only by the model's own authors, no independent check
How to set it up — step by step
- 1Open the model page at huggingface.co/Scantist/GLM-5.3-SLIM-GGUF and read the hardware requirements.
- 2Check whether you have a Mac Studio with 192 GB+ memory or an NVIDIA B200 GPU — without one, the model won't run.
- 3Send your agent the text from the block below so it explains the difference from the original model and the hardware you need, in plain words.
- 4If your hardware qualifies, ask the agent to help install llama.cpp and download the model files per the page's instructions.
- 5Start the model's server following the instructions and send a test request to confirm it works.
- 6Compare the model's answers to your own task — keep in mind only the authors measured its quality, not an independent source.
Text for your agent
Copy this and send it to your agent — Claude Code, Codex, any of them:
Open the GLM-5.3-SLIM-GGUF model card: https://huggingface.co/Scantist/GLM-5.3-SLIM-GGUF — explain in plain terms how the 192-expert version differs from the original 256-expert one, and what hardware it takes to run it.
You'd need a Mac Studio with 192 GB of memory or more, or a single NVIDIA B200 data-center GPU — this one won't run on home hardware.
Open on Hugging Face▌ More finds