Instructions to use WhirlwindAI/MetaNova-0.1-70M with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use WhirlwindAI/MetaNova-0.1-70M with Transformers:

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

pipe = pipeline("text-generation", model="WhirlwindAI/MetaNova-0.1-70M", trust_remote_code=True)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("WhirlwindAI/MetaNova-0.1-70M", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use WhirlwindAI/MetaNova-0.1-70M with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "WhirlwindAI/MetaNova-0.1-70M"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "WhirlwindAI/MetaNova-0.1-70M",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/WhirlwindAI/MetaNova-0.1-70M

SGLang

How to use WhirlwindAI/MetaNova-0.1-70M 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 "WhirlwindAI/MetaNova-0.1-70M" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "WhirlwindAI/MetaNova-0.1-70M",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

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 "WhirlwindAI/MetaNova-0.1-70M" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "WhirlwindAI/MetaNova-0.1-70M",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use WhirlwindAI/MetaNova-0.1-70M with Docker Model Runner:
```
docker model run hf.co/WhirlwindAI/MetaNova-0.1-70M
```

⚡ At a Glance

Model	Params	Final Avg	Highlights
WhirlwindAI/MetaNova-Test	70.55M	31.94%	🏅 ARC-Challenge (25.09%)
WhirlwindAI/MetaNova-0.1-70M	70.55M	35.36%	🏅 ArithMark-2 (27.12%)
SupraLabs/Supra-50M-Base	51.79M	38.60%	🏅 HellaSwag, ARC-Easy, PIQA, ARC Avg, Final Avg

🟣 Quick summary — The WhirlwindAI MetaNova models each have their own strengths: MetaNova-0.1-70M leads on ArithMark-2, while MetaNova-Test takes ARC-Challenge. Supra-50M-Base (external baseline) posts the highest Final Avg at 38.60% with the smallest parameter count.

🏁 Full Leaderboard

Benchmark	WhirlwindAI/ MetaNova-Test	WhirlwindAI/ MetaNova-0.1-BF16	SupraLabs/ Supra-50M-Base
Params	70.55M	70.55M	51.79M
HellaSwag	25.37%	27.40%	31.65%
ARC-Easy	26.89%	31.73%	45.58%
ARC-Challenge	25.09%	25.00%	24.66%
PIQA	52.29%	58.54%	61.53%
ArithMark-2	24.12%	27.12%	26.08%
ARC Avg	25.99%	28.37%	35.12%
Final Avg	31.94%	35.36%	38.60% 🏆

_{🏆 = best score in the row. Bold = row winner.}

📊 Metric-by-Metric View

🟣 HellaSwag

Model	Score
MetaNova-Test	25.37%	█████░░░░░░░
MetaNova-0.1-BF16	27.40%	█████░░░░░░░
Supra-50M-Base	31.65%	██████░░░░░░

🟣 ARC-Easy

Model	Score
MetaNova-Test	26.89%	█████░░░░░░░
MetaNova-0.1-BF16	31.73%	██████░░░░░░
Supra-50M-Base	45.58%	█████████░░░

🟣 ARC-Challenge

Model	Score
MetaNova-Test	25.09%	█████░░░░░░░
MetaNova-0.1-BF16	25.00%	█████░░░░░░░
Supra-50M-Base	24.66%	█████░░░░░░░

🟣 PIQA

Model	Score
MetaNova-Test	52.29%	██████████░░
MetaNova-0.1-BF16	58.54%	████████████
Supra-50M-Base	61.53%	████████████

🟣 ArithMark-2

Model	Score
MetaNova-Test	24.12%	█████░░░░░░░
MetaNova-0.1-BF16	27.12%	█████░░░░░░░
Supra-50M-Base	26.08%	█████░░░░░░░

🟣 Averages

Metric	MetaNova-Test	MetaNova-0.1-BF16	Supra-50M-Base
ARC Avg	25.99%	28.37%	35.12%
Final Avg	31.94%	35.36%	38.60%

🧠 THINKING Chat Format

Thinking Mode Non-Thinking Mode

Thinking Mode	Non-Thinking Mode
<\|im_start\|>user {query} /think<\|im_end\|> <\|im_start\|>assistant <think> {thinking_content} </think> {response}<\|im_end\|>	<\|im_start\|>user {query} /no_think<\|im_end\|> <\|im_start\|>assistant <think> </think> {response}<\|im_end\|> `</td> </tr>`

<|im_start|>user
{query} /think<|im_end|>
<|im_start|>assistant
<think>
{thinking_content}
</think>

{response}<|im_end|>

<|im_start|>user
{query} /no_think<|im_end|>
<|im_start|>assistant
<think>

</think>

{response}<|im_end|>

  </td>
</tr>

🔵 blue = user / assistant content • 🔴 red = mode switch + reasoning block

🧾 Run Summary


Model	WhirlwindAI/MetaNova-0.1-70M
Model size	70.5M params
Tensor type	BF16
Downloads last month	16
Models compared	3
Benchmarks	HellaSwag · ARC-Easy · ARC-Challenge · PIQA · ArithMark-2 · ARC Avg · Final Avg
Parameter range	51.79M – 70.55M
Final Avg range	31.94% – 38.60%
Top Final Avg	SupraLabs/Supra-50M-Base — 38.60%

_{⭐ If this model helped you, consider leaving a like on the Hub.}

_{WhirlwindAI · 2026}

Downloads last month: 243

Safetensors

Model size

70.5M params

Tensor type

BF16