Instructions to use nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 with Transformers:

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

pipe = pipeline("text-generation", model="nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-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-Nano-30B-A3B-BF16", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16", trust_remote_code=True)
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
Local Apps Settings

vLLM

How to use nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-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-Nano-30B-A3B-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-Nano-30B-A3B-BF16",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16

SGLang

How to use nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-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-Nano-30B-A3B-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-Nano-30B-A3B-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-Nano-30B-A3B-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-Nano-30B-A3B-BF16",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 with Docker Model Runner:
```
docker model run hf.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16
```

[Research] Adaptive-K Routing Validation: 33% Compute Savings on Nemotron 3 Nano

#41

by Gabrobals - opened Jan 19

Discussion

Gabrobals

Jan 19

Hi NVIDIA team! 👋

I've been working on Adaptive-K routing - an entropy-guided method for dynamic expert selection in MoE models. Today I validated it on Nemotron 3 Nano and wanted to share the results.

TL;DR

33.3% compute savings by dynamically selecting K∈{2,4,6} based on router entropy
Zero retraining required - inference-time only
Entropy-based selection: confident tokens use fewer experts

Results

Test Case	Router Entropy	Effective K	Savings
Easy tokens	5.26 bits	4.1	32.4%
Code tokens	5.28 bits	4.0	33.3%
Hard tokens	5.16 bits	3.9	34.4%
Average	5.23 bits	4.0	33.3%

The Insight

Nemotron 3's router entropy (measured via pre-top-k logits) averages 5.23 bits out of 7.0 max (log₂(128)). This means:

~75% of max entropy → router is moderately confident
Many tokens don't need all 6 experts
The shared expert provides a quality safety net for aggressive K reduction

Methodology

Since output_router_logits isn't supported, I used forward hooks on backbone.layers.X.mixer.gate to compute full 128-expert logits:

router_logits = hidden_states @ module.weight.T  # [batch, seq, 128]
probs = softmax(router_logits)
entropy = -sum(probs * log(probs))  # Per-token entropy


Why This Matters for Nemotron 3
Amplifies reasoning budget control: Users already control reasoning tokens - Adaptive-K automates compute optimization per-token
Shared expert synergy: The always-active shared expert means quality is maintained even at K=2
No retraining: Drop-in replacement for the router
Open Source
Full validation: https://github.com/Gabrobals/sbm-efficient
Results JSON: nemotron3_nano_validation.json
Validation script: nemotron3_entropy_validation.py
Questions for NVIDIA
Would you be interested in integrating Adaptive-K as an optional routing mode?
Is there a preferred way to contribute to the Nemotron cookbooks?
Any plans to expose output_router_logits in future versions?
Happy to collaborate on benchmarks or provide a PR to the NeMo repository!

Gabriele Balsamo
GitHub: @Gabrobals

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