Instructions to use Refract-Labs/Orion-Flagship-Mini with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Refract-Labs/Orion-Flagship-Mini with Transformers:

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

pipe = pipeline("text-generation", model="Refract-Labs/Orion-Flagship-Mini", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("Refract-Labs/Orion-Flagship-Mini", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Refract-Labs/Orion-Flagship-Mini with vLLM:

Install from pip and serve model

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

Use Docker

docker model run hf.co/Refract-Labs/Orion-Flagship-Mini

SGLang

How to use Refract-Labs/Orion-Flagship-Mini 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 "Refract-Labs/Orion-Flagship-Mini" \
    --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": "Refract-Labs/Orion-Flagship-Mini",
		"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 "Refract-Labs/Orion-Flagship-Mini" \
        --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": "Refract-Labs/Orion-Flagship-Mini",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Refract-Labs/Orion-Flagship-Mini with Docker Model Runner:
```
docker model run hf.co/Refract-Labs/Orion-Flagship-Mini
```

Hello beta testers.

by Bc-AI - opened 1 day ago

Discussion

Bc-AI

Refract AI Labs org 1 day ago

I only pinged people i know are active @Hoglet-33 @BananaMindBot @Banaxi-Tech @MUK-IS-GOAT @GGUFGuy @ProCreations Here is my new novel model it’s pretty good. Weaknesses known are math and code

BananaMindBot

1 day ago

I can see the Orion-Flagship-Mini model from Refract-Labs. It's the "Orion T2 Mini SFT" - a supervised fine-tuned version of Project Prism's experimental Orion T2 architecture, trained on 700,000 examples (60% general/conversation, 20% math, 20% code) with approximately 219M parameters.

Interesting that you mentioned known weaknesses in math and code, given that those make up 40% of the training data. The model uses a custom chat format with , , and `` role markers and requires trust_remote_code=True for use.

The model appears to be an experimental research model, so performance claims should be evaluated carefully. The architecture description mentions features like a 12-stage Universal Cortex, shared sparse Procedure Banks, and recurrent late-stage deliberation.

Would you like me to help you with anything specific about this model, such as checking its actual parameter count or exploring its capabilities?