Instructions to use nvidia/Qwen2.5-CascadeRL-RM-72B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use nvidia/Qwen2.5-CascadeRL-RM-72B with Transformers:

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

pipe = pipeline("text-generation", model="nvidia/Qwen2.5-CascadeRL-RM-72B")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForSequenceClassification

tokenizer = AutoTokenizer.from_pretrained("nvidia/Qwen2.5-CascadeRL-RM-72B")
model = AutoModelForSequenceClassification.from_pretrained("nvidia/Qwen2.5-CascadeRL-RM-72B")
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/Qwen2.5-CascadeRL-RM-72B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "nvidia/Qwen2.5-CascadeRL-RM-72B"
# 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/Qwen2.5-CascadeRL-RM-72B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/nvidia/Qwen2.5-CascadeRL-RM-72B

SGLang

How to use nvidia/Qwen2.5-CascadeRL-RM-72B 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/Qwen2.5-CascadeRL-RM-72B" \
    --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/Qwen2.5-CascadeRL-RM-72B",
		"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/Qwen2.5-CascadeRL-RM-72B" \
        --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/Qwen2.5-CascadeRL-RM-72B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use nvidia/Qwen2.5-CascadeRL-RM-72B with Docker Model Runner:
```
docker model run hf.co/nvidia/Qwen2.5-CascadeRL-RM-72B
```

Qwen2.5-CascadeRL-RM-72B

Description

Qwen2.5-CascadeRL-RM-72B is a reward model that is initialized with Qwen2.5-72B-Instruct and is fine-tuned using the Bradley-Terry objective to predict the human preference of LLM generation. It is used in Reinforcement Learning from Human Feedback (RLHF) stage in Nemotron-Cascade model family: Nemotron-Cascade-8B, Nemotron-Cascade-8B-Thinking, and Nemotron-Cascade-14B-Thinking.

Given a conversation between a human and an assistant, the reward model will give a human preference score for the final assistant turn.

For the training details, please refer to the technical report.

Usage Recommendations

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "nvidia/Qwen2.5-CascadeRL-RM-72B"
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    device_map="auto",  
    low_cpu_mem_usage=True,
    torch_dtype=torch.bfloat16,
)
tokenizer = AutoTokenizer.from_pretrained(model_name)
model.eval()

prompt = "Hello! How are you?"
response = "I am fine! Thanks for asking. How are you?"
messages = [{"role":"user","content":prompt}, {"role":"assistant","content":response}]
batch = tokenizer.apply_chat_template(
    messages, tokenize=True, add_generation_prompt=False,
    return_tensors="pt", return_dict=True
)
with torch.inference_mode():
    out = model(**batch, use_cache=False)
print(out.logits[0, -1, 0].item())

RewardBench

Model	Overall	Chat	Chat Hard	Safety	Reasoning
Qwen2.5-CascadeRL-RM-72B	95.15	98.60	89.69	93.92	98.40

Release Date

Dec 31, 2025

License

Your use of this model is governed by the NVIDIA Open Model License.

Citation

@article{Nemotron_Cascade_Scaling_Cascaded_Reinforcement_Learning,
  title={Nemotron-Cascade: Scaling Cascaded Reinforcement Learning for General-Purpose Reasoning Models},
  author={Wang, Boxin and Lee, Chankyu and Lee, Nayeon and Lin, Sheng-Chieh and Dai, Wenliang and Chen, Yang and Chen, Yangyi and Yang, Zhuolin and Liu, Zihan and Shoeybi, Mohammad and Catanzaro, Bryan and Ping, Wei},
  year={2025}
}

Downloads last month: 2,081

Safetensors

Model size

71B params

Tensor type

BF16

Collection including nvidia/Qwen2.5-CascadeRL-RM-72B

Nemotron-Cascade

Collection

Scaling Cascaded Reinforcement Learning for General-Purpose Reasoning Models • 14 items • Updated 7 days ago • 54

Paper for nvidia/Qwen2.5-CascadeRL-RM-72B

Nemotron-Cascade: Scaling Cascaded Reinforcement Learning for General-Purpose Reasoning Models

Paper • 2512.13607 • Published Dec 15, 2025 • 38