Instructions to use lukealonso/GLM-5.2-NVFP4 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use lukealonso/GLM-5.2-NVFP4 with Transformers:

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

pipe = pipeline("text-generation", model="lukealonso/GLM-5.2-NVFP4")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("lukealonso/GLM-5.2-NVFP4")
model = AutoModelForCausalLM.from_pretrained("lukealonso/GLM-5.2-NVFP4")
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]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use lukealonso/GLM-5.2-NVFP4 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "lukealonso/GLM-5.2-NVFP4"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "lukealonso/GLM-5.2-NVFP4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/lukealonso/GLM-5.2-NVFP4

SGLang

How to use lukealonso/GLM-5.2-NVFP4 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 "lukealonso/GLM-5.2-NVFP4" \
    --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": "lukealonso/GLM-5.2-NVFP4",
		"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 "lukealonso/GLM-5.2-NVFP4" \
        --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": "lukealonso/GLM-5.2-NVFP4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use lukealonso/GLM-5.2-NVFP4 with Docker Model Runner:
```
docker model run hf.co/lukealonso/GLM-5.2-NVFP4
```

Have you tested the AIME-2026 scores for this version?

by ghostplant - opened 23 days ago

Discussion

ghostplant

23 days ago

•

edited 23 days ago

Hi,

I evaluated the AIME-2026 benchmark using this NVFP4 model and only got 90%, whereas the officially reported score is 99%.

Under the same settings on GPU-5.1-NVFP4 (temperature=1, top_p=0.95), the result looks much better: 96% on AIME-26.

So I’m wondering whether this quantization may be suboptimal.

I didn’t (and currently can’t) evaluate the official BF16 weights because I don’t have enough disk space or GPU memory to run them.

FabianHeller

4 days ago

Just for comparison, I did run the AIME-2026 benchmark of nvidia/GLM-5.2-NVFP4 and also got only 90.83% with this config:

model: glm-5.2
api: custom
api_key_env: OPENAI_API_KEY
base_url: http://<HOST>:<PORT>/v1
max_tokens: 131072
concurrent_requests: 8
temperature: 1.0
top_p: 0.95
reasoning_effort: max
extra_body:
  thinking:
    type: enabled
read_cost: 0
write_cost: 0
human_readable_id: GLM 5.2 Nvidia NVFP4
date: "2026-06-22"
other_params:
  open: true
  creator: Z.ai
  parameters: 753
  active_parameters: 40
  huggingface_id: nvidia/GLM-5.2-NVFP4
stream_openai_chat_completions: true

and this command (delete the "vllm" dependency before from pyproject.toml):

uv run python scripts/run.py --comp aime/aime_2026 --models glm/glm-52-nvfp4 --n 4

The following problems failed:
Problem 15 and 30: all 4 runs failed, Problem 29: 2 runs failed, Problem 20: 1 run failed, Problem 29: 2 runs failed.

Maybe later I can try the failed problems with the Z.ai API to see if it is really better.

ghostplant

4 days ago

•

edited 3 days ago

Yes, I previously found that one of 3 wrong cases was due to incorrect match by the judger, so the correct scores should be around 93%.
If enabling tool_use, the scores go up to 99% for nvfp4.

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