Instructions to use nvidia/OpenMath-Nemotron-1.5B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use nvidia/OpenMath-Nemotron-1.5B with Transformers:

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

pipe = pipeline("text-generation", model="nvidia/OpenMath-Nemotron-1.5B")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("nvidia/OpenMath-Nemotron-1.5B")
model = AutoModelForCausalLM.from_pretrained("nvidia/OpenMath-Nemotron-1.5B")
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/OpenMath-Nemotron-1.5B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "nvidia/OpenMath-Nemotron-1.5B"
# 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/OpenMath-Nemotron-1.5B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/nvidia/OpenMath-Nemotron-1.5B

SGLang

How to use nvidia/OpenMath-Nemotron-1.5B 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/OpenMath-Nemotron-1.5B" \
    --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/OpenMath-Nemotron-1.5B",
		"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/OpenMath-Nemotron-1.5B" \
        --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/OpenMath-Nemotron-1.5B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use nvidia/OpenMath-Nemotron-1.5B with Docker Model Runner:
```
docker model run hf.co/nvidia/OpenMath-Nemotron-1.5B
```

Table results vs chart

by rawsh - opened Apr 23, 2025

Discussion

rawsh

Apr 23, 2025

•

edited Apr 23, 2025

Hi! Thanks for the great work. Curious for the 1.5B, the pass@1 numbers in the table seem inconsistent with the chart and results in the paper:

Model	AIME24	AIME25	HMMT-24-25	HLE-Math
DeepSeek-R1-Distill-Qwen-1.5B	26.8 (60.0)	21.4 (36.7)	14.2 (26.5)	2.9 (5.0)
OpenMath-Nemotron-1.5B CoT	61.6 (80.0)	49.5 (66.7)	39.9 (53.6)	5.4 (5.4)
OpenMath-Nemotron-1.5B TIR	52.0 (83.3)	39.7 (70.0)	37.2 (60.7)	2.5 (6.2)
+ Self GenSelect	83.3	70.0	62.2	7.9
+ 32B GenSelect	83.3	70.0	62.8	8.3

What are the details of the 61.6% and 80% results? The 43% pass@1 I'm eyeballing from the figure matches DeepScaleR closely, but 61.6% pass@1 would be absurdly impressive...

majority@64 = 58.2%:

eyeballing pass@1 = ~41%, majority@64 = ~55% (first sft stage)

eyeballing pass@1 = ~43%, majority@64 = ~58%

Thank you very much for your time!

igitman

NVIDIA org Apr 24, 2025

Hi @rawsh , thanks for your interest in our work! Please note that most of our results are presented on the comp-math-24-25 evaluation set that is a combination of 60 AIME problems and 196 HMMT problems from 2024 and 2025. Table 7 has a breakdown of those numbers into AIME24, AIME25 and HMMT sets. So the 61.6% and 80% are the results our 1.5B model gets on AIME24 with pass@1 and majority@64 correspondingly. While the other numbers you're referring to are on the combined set, which is generally harder and thus has lower scores.

Hope this clears it up and let us know if you have any more questions. Please also keep in mind that even though we run 64 generations, there is still a lot of variance and it's easy to see +=3% in AIME scores as those only have 30 problems.

rawsh

Apr 24, 2025

•

edited Apr 24, 2025

@igitman Thank you for the explanation! Congrats on such strong models + pipeline!

I was able to replicate AIME24 with lighteval:

Test Set	Average Accuracy@64	Standard Error
AIME 2024	0.6099	0.0692

avg@16:

Test Set	Average Accuracy@16	Standard Error
AIME 2024	0.6521	0.0703
AMC 22-23	0.8464	0.0318
AMC 24	0.7319	0.0572

rawsh changed discussion status to closed Apr 27, 2025

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