Instructions to use SteelStorage/Umbra-v3-MoE-4x11b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use SteelStorage/Umbra-v3-MoE-4x11b with Transformers:

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

pipe = pipeline("text-generation", model="SteelStorage/Umbra-v3-MoE-4x11b")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("SteelStorage/Umbra-v3-MoE-4x11b")
model = AutoModelForCausalLM.from_pretrained("SteelStorage/Umbra-v3-MoE-4x11b")
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

vLLM

How to use SteelStorage/Umbra-v3-MoE-4x11b with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "SteelStorage/Umbra-v3-MoE-4x11b"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "SteelStorage/Umbra-v3-MoE-4x11b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/SteelStorage/Umbra-v3-MoE-4x11b

SGLang

How to use SteelStorage/Umbra-v3-MoE-4x11b 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 "SteelStorage/Umbra-v3-MoE-4x11b" \
    --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": "SteelStorage/Umbra-v3-MoE-4x11b",
		"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 "SteelStorage/Umbra-v3-MoE-4x11b" \
        --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": "SteelStorage/Umbra-v3-MoE-4x11b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use SteelStorage/Umbra-v3-MoE-4x11b with Docker Model Runner:
```
docker model run hf.co/SteelStorage/Umbra-v3-MoE-4x11b
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

Creator: SteelSkull

About Umbra-v3-MoE-4x11b: A Mixture of Experts model designed for general assistance with a special knack for storytelling and RP/ERP

Integrates models from notable sources for enhanced performance in diverse tasks.

Source Models:

Update-Log:

The [Umbra Series] keeps rolling out from the [Lumosia Series] garage, aiming to be your digital Alfred with a side of Shakespeare for those RP/ERP nights.

What's Fresh in v3?

Didn’t reinvent the wheel, just slapped on some fancier rims. Upgraded the models and tweaked the prompts a bit. Now, Umbra's not just a general use LLM; it's also focused on spinning stories and "Stories".

Negative Prompt Minimalism

Got the prompts to do a bit of a diet and gym routine—more beef on the positives, trimming down the negatives as usual with a dash of my midnight musings.

Still Guessing, Aren’t We?

Just so we're clear, "v3" is not the messiah of updates. It’s another experiment in the saga.

Dive into Umbra v3 and toss your two cents my way. Your feedback is the caffeine in my code marathon.

Exl2 available by:

EXL2-Rpcal = AzureBlack

GGUF = mradermacher

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	73.09
AI2 Reasoning Challenge (25-Shot)	68.43
HellaSwag (10-Shot)	87.83
MMLU (5-Shot)	65.99
TruthfulQA (0-shot)	69.30
Winogrande (5-shot)	83.90
GSM8k (5-shot)	63.08

Downloads last month: 66

Safetensors

Model size

36B params

Tensor type

BF16

Model tree for SteelStorage/Umbra-v3-MoE-4x11b

Himitsui/Kaiju-11B

Sao10K/Fimbulvetr-11B-v2

beberik/Nyxene-v3-11B

decapoda-research/Antares-11b-v2

Merge model

this model

Quantizations

2 models

Evaluation results

normalized accuracy on AI2 Reasoning Challenge (25-Shot)
test set Open LLM Leaderboard

68.430
normalized accuracy on HellaSwag (10-Shot)
validation set Open LLM Leaderboard

87.830
accuracy on MMLU (5-Shot)
test set Open LLM Leaderboard

65.990
mc2 on TruthfulQA (0-shot)
validation set Open LLM Leaderboard

69.300
accuracy on Winogrande (5-shot)
validation set Open LLM Leaderboard

83.900
accuracy on GSM8k (5-shot)
test set Open LLM Leaderboard

63.080