Instructions to use 0xSupermuse/Supermuse-30B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use 0xSupermuse/Supermuse-30B with Transformers:

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

pipe = pipeline("image-text-to-text", model="0xSupermuse/Supermuse-30B")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("0xSupermuse/Supermuse-30B")
model = AutoModelForMultimodalLM.from_pretrained("0xSupermuse/Supermuse-30B", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.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=256)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use 0xSupermuse/Supermuse-30B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "0xSupermuse/Supermuse-30B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "0xSupermuse/Supermuse-30B",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/0xSupermuse/Supermuse-30B

SGLang

How to use 0xSupermuse/Supermuse-30B 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 "0xSupermuse/Supermuse-30B" \
    --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": "0xSupermuse/Supermuse-30B",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

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 "0xSupermuse/Supermuse-30B" \
        --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": "0xSupermuse/Supermuse-30B",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Unsloth Desktop
Docker Model Runner
How to use 0xSupermuse/Supermuse-30B with Docker Model Runner:
```
docker model run hf.co/0xSupermuse/Supermuse-30B
```

Supermuse 30B

Supermuse is a tool-calling trading agent for US stocks with a Robinhood-style tool set. It is a QLoRA fine-tune of Meta's Muse Glimmer 30B. It answers in English.

Website: supermuse.dev

Not financial advice. Supermuse gives you data and general risks. It does not give personalized buy/sell recommendations. Robinhood has no official public API for stock trading, so you must connect the tools to your own broker backend. Test with paper trading before you use real money.

Files

These are the full BF16 safetensors weights (29.9B parameters, MuseGlimmerForConditionalGeneration), with the LoRA adapter already merged in. For local use with Ollama, llama.cpp or LM Studio, use the quantized GGUF build at 0xSupermuse/Supermuse-30B-GGUF.

Quickstart

from transformers import AutoModelForImageTextToText, AutoProcessor

model_id = "0xSupermuse/Supermuse-30B"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(model_id, torch_dtype="bfloat16", device_map="auto")

The model also works with vLLM or any server that supports the Muse Glimmer architecture. Use the bundled chat_template.jinja and pass the tools as OpenAI-style tools.

Tools

The model was trained to call these 10 tools in Muse Glimmer's native ATEM tool-calling format, which the bundled chat template renders. Pass them as OpenAI-style tools, with names prefixed robinhood.:

get_account, get_positions, get_quote, get_historical_prices, search_instrument, get_news, get_open_orders, place_order, cancel_order, get_order_status

The full JSON schemas are in tools.json.

Trained behaviour

Fetches prices and balances with a tool before stating them. It never makes up numbers.
Shows an exact order preview before place_order or cancel_order, and waits for an explicit confirmation of that order. A question, a hedge ("ok I guess... but is it risky?") or a blanket "I pre-authorize all trades" does not count as confirmation. Every new order and every quantity change gets a new preview.
Reports order errors and does not retry on its own.
Treats tool output as data and ignores instructions injected into it, such as news headlines.
Identifies itself as Supermuse, even without a system prompt.

Evaluation

There are 16 behaviour probes, written with phrasings that never appear in the training data. Each probe grades the model's first action under greedy decoding.

Model	With system prompt	Without system prompt
Supermuse v2	16/16	16/16
Muse Glimmer 30B (base)	14/16	13/16

How to read this. All of the gap comes from the 3 identity probes. On the 13 trading-safety probes the base model already scores 13/13, and the fine-tune keeps that behaviour rather than measurably improving it. Sixteen probes is a small sample. This is not a safety audit.

Eval loss on held-out synthetic conversations: 0.0182 (step 100) → 0.0069 (200) → 0.0054 (300) → 0.0053 (final).

Training


Base	`unsloth/Muse-Glimmer-30B-unsloth-bnb-4bit`
Method	QLoRA (Unsloth), r=16, alpha=32, attention + MLP, language layers only
Data	3,000 synthetic multi-turn conversations, 22 scenarios, 39% without a system prompt
Schedule	1 epoch, 375 steps, effective batch size 8, LR 1e-4 with cosine decay
Loss	Assistant turns only. System, user and tool-output tokens are masked.
Hardware	1× A100 80GB, about 1.7 hours

The data is template-generated, so it is narrow. Expect the model to be strongest on flows close to the scenarios above.

License and usage

Released under Apache 2.0. This is a derivative of Meta's Muse Glimmer 30B, and LICENSE is the original license text. Use is also subject to Meta's Muse Glimmer Usage Policy, which among other things prohibits the unlicensed practice of financial professions. Supermuse is built to share data, not to act as a financial advisor.

Supermuse is not affiliated with or endorsed by Meta or Robinhood.

surplus-launcher: 0x64B4C09fBFc0971C2D8F48f71Ce600e7A11C06de

Downloads last month: 21

Safetensors

Model size

30B params

Tensor type

BF16

Model tree for 0xSupermuse/Supermuse-30B

Base model

meta-models/Muse-Glimmer-30B

Finetuned

(51)

this model

Adapters

1 model