Instructions to use nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16 with Transformers:

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

pipe = pipeline("text-generation", model="nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16", trust_remote_code=True, device_map="auto")
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
Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16

SGLang

How to use nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16 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 "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16" \
    --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": "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16",
		"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 "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16" \
        --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": "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16 with Docker Model Runner:
```
docker model run hf.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16
```

Disobedient and "rude"

#16

by deleted - opened Mar 15

Discussion

deleted

Mar 15

The model is excellent in reasoning and such, but it's quite "disobedient" and even (in some sense) "rude", in regards to modifications to its instructions given during agentic coding.

For example, if you kick off a task, it can work through that. If it stops and asks something, and you respond, it can work through that appropriately (usually).

BUT, if you interrupt its reasoning, and tell it that it should do something else, or that it's going off course and needs to follow the original instruction, it often just continues on, giving little or not regard to the additional instruction.

AT LEAST, instead of being ignored, instead of output of the form "<further_output"> I would expect "OK, I hear you, I'll incorporate that as I go forward. ", or (if it must) "The user says . Actually, I think that's incorrect due to ; I'm still fully aligned with everything the user has requested, and so I can to proceed as before. In this framework, the user can simply interrupt me again if I've gotten this wrong, so I'll proceed. ".

deleted

Mar 29

•

edited Mar 29

Actually, it's disobedient from the outset, too. For example, if I tell it to make a skill, and to search online for the format of skill.md files, it will decide to initially search the filesystem for skill.md files instead. Not ideal.

deleted

Apr 10

•

edited Apr 10

To clarify: I believe there's a gap in the training dataset here: it's likely trained on a simple back and forth flow, whereas the training should include interruptions for questions, additions, course corrections (of the model's CoT), and "changes of mind" (of the user) and so on.

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