Instructions to use CraneAILabs/crane-nemotron-nano-4b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use CraneAILabs/crane-nemotron-nano-4b with Transformers:

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

pipe = pipeline("text-generation", model="CraneAILabs/crane-nemotron-nano-4b", 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("CraneAILabs/crane-nemotron-nano-4b", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("CraneAILabs/crane-nemotron-nano-4b", trust_remote_code=True, device_map="auto")
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 Settings

vLLM

How to use CraneAILabs/crane-nemotron-nano-4b with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "CraneAILabs/crane-nemotron-nano-4b"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "CraneAILabs/crane-nemotron-nano-4b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/CraneAILabs/crane-nemotron-nano-4b

SGLang

How to use CraneAILabs/crane-nemotron-nano-4b 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 "CraneAILabs/crane-nemotron-nano-4b" \
    --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": "CraneAILabs/crane-nemotron-nano-4b",
		"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 "CraneAILabs/crane-nemotron-nano-4b" \
        --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": "CraneAILabs/crane-nemotron-nano-4b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use CraneAILabs/crane-nemotron-nano-4b with Docker Model Runner:
```
docker model run hf.co/CraneAILabs/crane-nemotron-nano-4b
```

Crane Nemotron 4B

An experimental hybrid-architecture model for East African languages and agentic tool-calling.

Crane Nemotron 4B is a research model from Crane AI Labs exploring hybrid Mamba + Transformer architectures for multilingual and agentic use. Most open language models are pure Transformers. This one is not: it is built on NVIDIA's Nemotron-Nano (NemotronH) backbone, which interleaves Mamba-2 state-space layers with attention layers. We were curious how well a hybrid SSM/attention model absorbs low-resource languages and structured tool use, so we adapted it to seven East African languages, taught it agentic tool-calling, and used reinforcement learning to correct multilingual drift and tool-triggering.

This is an experiment, shared in that spirit. The numbers and behavior below are what we measured, including where the model still falls short.

Highlights

Hybrid architecture (Mamba-2 state-space + attention), not a pure Transformer.
7 East African languages: Amharic, English, Hausa, Kinyarwanda, Luganda, Somali, Swahili.
Agentic tool-calling: web search, weather, and calculator, with a 90% tool-trigger decision accuracy.
Large reasoning and knowledge gains over the base checkpoint after continued pretraining, unified SFT, and RL.

Strengths (measured)

Benchmark	Base	Crane Nemotron 4B
MMLU (knowledge)	29.8	53.8
GSM8K (math reasoning)	60.4	75.2
Tool-trigger accuracy (80 cases)	~52 (SFT)	90
Language-consistency reward (RL)	0.73	0.90
FLORES X to English chrF		Swahili 49 / Luganda 36

Coherent multilingual chat and persona across the seven languages.
Tool over-calling eliminated; tool under-calling reduced to 6/40 on the balanced eval.
Translation into English is strong for the higher-resource languages (Swahili, Luganda).

Limitations

We prefer to be direct about what does not work yet.

AfriMMLU is near chance (~23%): multilingual knowledge QA remains weak.
English-to-X generation is weaker than X-to-English.
Amharic generation (Ge'ez script) is currently unreliable.
Calculator tool still occasionally narrates ("let me use the calculator...") instead of emitting the call on harder expressions (residual ~6/40 under-call).
This is a research checkpoint, not a production-hardened model.

Architecture and training

Base: NVIDIA Nemotron-Nano 4B (NemotronH), a hybrid Mamba-2 + Transformer model.
Pipeline: continued pretraining on East African language data, unified capability SFT (chat + persona + tool-calling in one pass so nothing washes out), then GRPO reinforcement learning for language consistency and tool-triggering.
Tool-calling uses the qwen3_xml tool-call format and supports OpenAI-style tools with tool_choice="auto".

Intended use

Research and experimentation with hybrid architectures for African-language and agentic applications. Not intended for high-stakes or unsupervised production use.

Acknowledgements

Built on NVIDIA Nemotron-Nano. Developed by Crane AI Labs as part of ongoing work on East African language models and hybrid architectures.

Downloads last month: 284

Safetensors

Model size

4B params

Tensor type

BF16