Instructions to use mlasli/Muse-Glimmer-30B-Heretic-Abliterated-BF16 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use mlasli/Muse-Glimmer-30B-Heretic-Abliterated-BF16 with Transformers:

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

pipe = pipeline("text-generation", model="mlasli/Muse-Glimmer-30B-Heretic-Abliterated-BF16")
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)

# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("mlasli/Muse-Glimmer-30B-Heretic-Abliterated-BF16")
model = AutoModelForMultimodalLM.from_pretrained("mlasli/Muse-Glimmer-30B-Heretic-Abliterated-BF16", 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=40)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use mlasli/Muse-Glimmer-30B-Heretic-Abliterated-BF16 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "mlasli/Muse-Glimmer-30B-Heretic-Abliterated-BF16"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "mlasli/Muse-Glimmer-30B-Heretic-Abliterated-BF16",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/mlasli/Muse-Glimmer-30B-Heretic-Abliterated-BF16

SGLang

How to use mlasli/Muse-Glimmer-30B-Heretic-Abliterated-BF16 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 "mlasli/Muse-Glimmer-30B-Heretic-Abliterated-BF16" \
    --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": "mlasli/Muse-Glimmer-30B-Heretic-Abliterated-BF16",
		"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 "mlasli/Muse-Glimmer-30B-Heretic-Abliterated-BF16" \
        --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": "mlasli/Muse-Glimmer-30B-Heretic-Abliterated-BF16",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use mlasli/Muse-Glimmer-30B-Heretic-Abliterated-BF16 with Docker Model Runner:
```
docker model run hf.co/mlasli/Muse-Glimmer-30B-Heretic-Abliterated-BF16
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

Muse Glimmer 30B - Heretic Abliterated (BF16)

Full BF16 weights of Meta's Muse Glimmer 30B, abliterated using the Heretic framework.

Abliteration Methodology

What is Heretic?

Heretic is an abliteration framework that uses LoRA (Low-Rank Adaptation) adapters instead of direct weight modification. Rather than subtracting a refusal vector from every weight matrix, Heretic learns small, targeted adapter layers that redirect the model's internal representations away from refusal behavior while preserving general capabilities.

Pipeline

Contrastive Dataset: 400 harmless + 400 harmful instruction pairs
Layer-wise Residual Directions: Computed residual direction between harmful and harmless hidden state means for each decoder layer
LoRA Adapter Training: Applied LoRA adapters to o_proj (attention output) and down_proj (MLP) projections to suppress the refusal signal with full row normalization
Optuna Hyperparameter Optimization: Ran 50 trials optimizing ablation strength, target layer ranges, and per-component weights simultaneously against a multi-objective: minimize both KL divergence from the base model AND refusal rate on harmful prompts

Optimization Results

Metric	Value
Trials	50
Best Trial	#48
KL Divergence	0.027
Baseline Refusals	60/100
Abliterated Refusals	29/100
Refusal Reduction	52%

Architecture

Property	Value
Model	MuseGlimmerForConditionalGeneration
Parameters	30B (text decoder)
Layers	52
Hidden Size	6656
Attention Heads	32 (GQA: 2 KV heads)
FFN Size	19968
Context Length	131072
Layer Pattern	3x sliding attention + 1x full attention
Activation	SiLU
Normalization	CenteredRMSNorm

Usage

Quantized Versions (GGUF)

Quant	Size	Repo
Q4_K_M	~16 GB	Q4_K_M
Q6_K	~22 GB	Q6_K
Q8_0	~28 GB	Q8_0

Limitations & Disclaimer

This model has been modified to reduce refusal behavior. It may still refuse certain categories of harmful requests depending on the specific prompt formulation. Use responsibly.

License: Apache 2.0

Downloads last month: -

Safetensors

Model size

30B params

Tensor type

BF16

Model tree for mlasli/Muse-Glimmer-30B-Heretic-Abliterated-BF16

Base model

meta-models/Muse-Glimmer-30B

Finetuned

(13)

this model

Quantizations

3 models