Instructions to use nvidia/Nemotron-Cascade-14B-Thinking with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use nvidia/Nemotron-Cascade-14B-Thinking with Transformers:

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

pipe = pipeline("text-generation", model="nvidia/Nemotron-Cascade-14B-Thinking")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("nvidia/Nemotron-Cascade-14B-Thinking")
model = AutoModelForCausalLM.from_pretrained("nvidia/Nemotron-Cascade-14B-Thinking")
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

vLLM

How to use nvidia/Nemotron-Cascade-14B-Thinking with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "nvidia/Nemotron-Cascade-14B-Thinking"
# 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-14B-Thinking",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/nvidia/Nemotron-Cascade-14B-Thinking

SGLang

How to use nvidia/Nemotron-Cascade-14B-Thinking 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-14B-Thinking" \
    --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-14B-Thinking",
		"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-14B-Thinking" \
        --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-14B-Thinking",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use nvidia/Nemotron-Cascade-14B-Thinking with Docker Model Runner:
```
docker model run hf.co/nvidia/Nemotron-Cascade-14B-Thinking
```

Extremely misleading benchmarks

by rombodawg - opened Dec 16, 2025

Discussion

rombodawg

Dec 16, 2025

•

edited Dec 16, 2025

The benchmark photo you have on your model card uses almost all diffrent models which is extremely misleading. Making people think your model is better than it really is.

For example qwen3-next-80b is only on the first benchmark. Why not the other 2??? Is it that you are trying to hide how bad your model really is?

ianncity

Dec 19, 2025

14b model vs 80b model, I do think they picked the best looking benchmarks obviously; but outperforming a model that's magnitudes larger in parameters on any benchmark is impressive

rombodawg

Dec 19, 2025

14b model vs 80b model, I do think they picked the best looking benchmarks obviously; but outperforming a model that's magnitudes larger in parameters on any benchmark is impressive

Classic Nvidia, maybe next time they'll have chart with 4x Token-Gen based on the first token. (Frame gen joke)

urroxyz

Dec 19, 2025

This comment has been hidden (marked as Resolved)

spanspek

Jan 25

This is a general problem with model releases and model maker showing selective benchmarks

The model makers should rather do this:

Select comparable models
Get the same metrics for all of the models selected in 1
Show the scores for each metric, for each selected model

I'm also of the opinion that even small models must include the current SOTA GPT, Claude, Gemini just to contexualize the model in the industry as a whole.

Or maybe do this as two sets of benchmarks:

Against peers as described above
Against industry-leading SOTA models

This way we can understand how the model measures up against it's immediate peers but also how it performs relative to the leading models out there

spanspek

Jan 25

That being said, I do find this model to be quite good with coding (for it's size)

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