Instructions to use locailabs/Juno-N-Coder-25B-A3B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use locailabs/Juno-N-Coder-25B-A3B with Transformers:

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

pipe = pipeline("text-generation", model="locailabs/Juno-N-Coder-25B-A3B", 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("locailabs/Juno-N-Coder-25B-A3B", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("locailabs/Juno-N-Coder-25B-A3B", 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 locailabs/Juno-N-Coder-25B-A3B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "locailabs/Juno-N-Coder-25B-A3B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "locailabs/Juno-N-Coder-25B-A3B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/locailabs/Juno-N-Coder-25B-A3B

SGLang

How to use locailabs/Juno-N-Coder-25B-A3B 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 "locailabs/Juno-N-Coder-25B-A3B" \
    --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": "locailabs/Juno-N-Coder-25B-A3B",
		"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 "locailabs/Juno-N-Coder-25B-A3B" \
        --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": "locailabs/Juno-N-Coder-25B-A3B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use locailabs/Juno-N-Coder-25B-A3B with Docker Model Runner:
```
docker model run hf.co/locailabs/Juno-N-Coder-25B-A3B
```

Juno-N-Coder-25B

Juno-N-Coder-25B is a coding-specialised derivative of NVIDIA Nemotron 3.5 Lightning 30B, produced by structured expert pruning with our SPACE algorithm (Specialisation Pruning Algorithm for Compression of Experts). This is the first model in the Juno series developed to run on-prem on the Locai One.

Locai Labs was one of NVIDIA's early-access partners for Nemotron 3.5 Lightning, and we want to thank the NVIDIA team for giving us early access to the model and supporting our work.

The goal with Juno-N-Coder was to create a coding-specialised version of Nemotron 3.5 Lightning purely through pruning and without any recovery fine-tuning. Using SPACE, we evaluated each expert's contribution to the target capability, in this case software development, removing the bottom 25% and renormalising the router so that the gate distribution remains correctly scaled.

The results are summarised in the model's performance profile below. Juno-N retains performance across four software engineering benchmarks, while degrading in general knowledge and factual recall as measured by MMLU Redux and SimpleQA.

Benchmark	Juno-N-Coder-25B	Nemotron 3.5 Lightning 30B
LiveCodeBench v6 (`test_v6_2408_2505`)	77.09	77.97
HumanEval+	90.55	89.33
MBPP+	80.29	81.49
SciCode (subtask)	36.98	36.09
MMLU Redux	82.18	90.00
SimpleQA (rubric)	37.88	47.95

MMLU Redux subject group	Juno-N-Coder-25B	Base	Δ
Maths & CS	94.6	94.9	-0.3
Physical sciences	88.3	94.4	-6.1
Economics	78.7	87.4	-8.7
Medicine	72.7	84.4	-11.7
World facts	71.7	84.0	-12.3
Humanities	75.9	88.5	-12.6

All models were evaluated using NVIDIA NeMo Evaluator at identical settings for both models: temperature 1.0, top_p 0.95, a 65,536-token generation limit, BF16 weights, and reasoning traces separated from the response before scoring.

Usage

vllm serve locailabs/Juno-N-Coder-25B \
    --trust-remote-code \
    --max-model-len 131072 \
    --reasoning-parser ultra_v3 \
    --reasoning-parser-plugin "$PARSER" \
    --tool-call-parser qwen3_coder \
    --enable-auto-tool-choice \
    --enable-prefix-caching

Citation

@misc{juno-n-coder-25b,
  title  = {Juno-N-Coder-25B: Specialisation Pruning for Compression of Experts},
  author = {Locai Labs},
  year   = {2026},
  url    = {https://huggingface.co/locailabs/Juno-N-Coder-25B}
}

Downloads last month: -

Safetensors

Model size

25B params

Tensor type

F32

BF16

Collection including locailabs/Juno-N-Coder-25B-A3B

Juno-N

Collection

Capability targeted compression of Nemotron 3.5 Lightning using our SPACE compression algorithm • 4 items • Updated about 4 hours ago