Instructions to use zai-org/GLM-5.2 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use zai-org/GLM-5.2 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="zai-org/GLM-5.2")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("zai-org/GLM-5.2")
model = AutoModelForCausalLM.from_pretrained("zai-org/GLM-5.2")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Inference
HuggingChat
Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use zai-org/GLM-5.2 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "zai-org/GLM-5.2"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "zai-org/GLM-5.2",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/zai-org/GLM-5.2

SGLang

How to use zai-org/GLM-5.2 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "zai-org/GLM-5.2" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "zai-org/GLM-5.2",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "zai-org/GLM-5.2" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "zai-org/GLM-5.2",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use zai-org/GLM-5.2 with Docker Model Runner:
```
docker model run hf.co/zai-org/GLM-5.2
```

Dense 60-120B model

#31

by rtzurtz - opened 18 days ago

Discussion

rtzurtz

18 days ago

•

edited 18 days ago

A 60B dense model may perform on par to a ~270-330B MoE model, if one considers sqrt(total B parameters * active B parameter), as real-life examples of dense vs MoE of similar size difference still confirm this formula.
And a 120B dense one would perform on par to an accordingly ~4.5-5.5 times larger MoE model.
But in each case, both would fit in so many more people's hardware. Even if it's a quant in all cases, it's still roughly a 5 times difference in memory requirement.

High RAM/memory prices, to the point where buying more RAM vs using a GPU, does not make sense any more.
A dense model provides a much better performance for a given parameter or memory requirement size.
For me and from comments, for many, waiting time can be much less relevant vs a high quality answer.
A 60B Q4_K_M quant (~40 GB + some context) will still fit in a 2 x 24 GB setup and run pretty fast.
(DDR6 and GDDR8 will speed up and compensate for such a model in the future.)

KeinNiemand

3 days ago

•

edited 3 days ago

I second this.
A modern 70B dense model with architecture, data and training methodology on par with something like GLM 5.2 would be a dream come true,
With my 42GB of VRAM, 70B is the Sweet spot, while I can also run 120B MoE, a good 70B dense should absolutely destroy a 120B MoE, while also running faster on my hardware.

OblxDev

3 days ago

With my 42GB of VRAM, 70B is the Sweet spot [...]

May I ask more about how you run GLM 5.2 within 42Gb vram ?
And what is the said hardware, software and outputs ?

This would greatly help 😄

KeinNiemand

3 days ago

•

edited 3 days ago

With my 42GB of VRAM, 70B is the Sweet spot [...]

May I ask more about how you run GLM 5.2 within 42Gb vram ?
And what is the said hardware, software and outputs ?

This would greatly help 😄

I can't run GLM 5.2 in 42GB of VRAM. I could run a 70B dense version or a 100-150B MoE (for the MoE I'm using ik_llama.cpp --n-cpu-moe to ofload experts to system ram)

OblxDev

3 days ago

I can't run GLM 5.2 in 42GB of VRAM. I could run a 70B dense version or a 100-150B MoE (for the MoE I'm using ik_llama.cpp --n-cpu-moe to ofload experts to system ram)

First, thanks for replying 👌
Really helpful to understand better how to use my gpus for more than "144Hz gaming + ollama". Indeed offloading looks to be the right path to follow.

And super great work you are sharing, i'll give CYOA a try !

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment