Instructions to use nvidia/Nemotron-Cascade-2-30B-A3B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use nvidia/Nemotron-Cascade-2-30B-A3B with Transformers:

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

pipe = pipeline("text-generation", model="nvidia/Nemotron-Cascade-2-30B-A3B", 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/Nemotron-Cascade-2-30B-A3B", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("nvidia/Nemotron-Cascade-2-30B-A3B", 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]:]))

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use nvidia/Nemotron-Cascade-2-30B-A3B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "nvidia/Nemotron-Cascade-2-30B-A3B"
# 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/Nemotron-Cascade-2-30B-A3B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/nvidia/Nemotron-Cascade-2-30B-A3B

SGLang

How to use nvidia/Nemotron-Cascade-2-30B-A3B 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/Nemotron-Cascade-2-30B-A3B" \
    --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/Nemotron-Cascade-2-30B-A3B",
		"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/Nemotron-Cascade-2-30B-A3B" \
        --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/Nemotron-Cascade-2-30B-A3B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use nvidia/Nemotron-Cascade-2-30B-A3B with Docker Model Runner:
```
docker model run hf.co/nvidia/Nemotron-Cascade-2-30B-A3B
```

MoE efficiency for agentic workflows: thinking mode trade-offs

#22

by O96a - opened Mar 29

Discussion

O96a

Mar 29

Gold medal in IMO 2025 and IOI 2025 is a strong signal — the 3B activated parameters from a 30B MoE is an attractive efficiency profile for edge deployment. I'm particularly interested in the thinking vs instruct mode trade-off.

In agentic pipelines (LangGraph, CrewAI), I've observed that models with explicit reasoning tokens often excel at planning but struggle with tool-calling consistency when the reasoning path is too verbose. The Nemotron-Cascade-2 benchmark shows strong LiveCodeBench (87.2) and BFCL v4 (52.9) — but there's a notable gap between code reasoning and function-calling performance.

Two practical questions:

How does thinking mode latency scale for multi-step agent workflows? If the model generates 500+ reasoning tokens before each tool call, the RTT for complex agents can become prohibitive.
For production deployment, has NVIDIA observed any degradation in tool-calling accuracy when switching between thinking and instruct modes within the same session? This is critical for agents that need both reasoning depth and structured output.

The MoE architecture with 3B activated is compelling — if reasoning efficiency holds, this could be a strong candidate for local agent deployments.

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