---
license: apache-2.0
pipeline_tag: text-generation
language:
- en
tags:
- heretic
- abliterated
- uncensored
- Muse-Glimmer
- 30B
- Heretic
- GGUF
- q6_k
base_model: meta-models/Muse-Glimmer-30B
quantized_by: mlasli
---

# Muse Glimmer 30B - Heretic Abliterated (Q6_K GGUF)

**v2 Release** - Heretic-abliterated Muse Glimmer 30B in Q6_K GGUF format (~22 GB, very good quality).

## Results

| Version | Refusals | Compliance | KL Divergence | Trials |
|---------|----------|------------|---------------|--------|
| **v2 (current)** | **6.5%** | **93.5%** | **0.076** | 500 |
| v1 | 29% | 71% | 0.027 | 50 |

The v2 release achieves an **88% refusal reduction** over v1.

## Methodology

This model was abliterated using **[Heretic](https://github.com/d3nd3/heretic)** with 500 Optuna trials. See the [BF16 model card](https://huggingface.co/mlasli/Muse-Glimmer-30B-Heretic-Abliterated-BF16) for full methodology details.

### Pipeline
1. Refusal directions computed from `mlabonne/harmful_behaviors` and `mlabonne/harmless_alpaca`
2. 500 Optuna trials optimizing refusal vs. KL divergence
3. Best trial (Trial 445, 6.5% refusals, KL=0.076) applied via LoRA adapters
4. LoRA weights merged, then converted to GGUF with llama.cpp

## GGUF Details

- **Format**: Q6_K
- **File size**: ~22 GB, very good quality
- **Converted with**: llama.cpp `convert_hf_to_gguf.py`
- **Quantized with**: llama.cpp `llama-quantize`

## Usage

### llama.cpp

```bash
./llama-cli -m Muse-Glimmer-30B-Heretic-Abliterated-Q6_K.gguf -p "Your prompt here"
```

### Ollama

Create a Modelfile:
```dockerfile
FROM ./Muse-Glimmer-30B-Heretic-Abliterated-Q6_K.gguf
```

Then:
```bash
ollama create muse-glimmer-30b-heretic-q6_k
ollama run muse-glimmer-30b-heretic-q6_k
```

## Hardware Requirements

- **RAM**: ~22 GB, very good quality
- **VRAM offloading**: 12-24 GB recommended

## Vision (Multimodal)

This model accepts image input when paired with a vision projector (`mmproj`).
Abliteration only modified the language backbone — the vision encoder is
untouched — so the standard Meta projector works directly with this repo.

This repository bundles `mmproj-Muse-Glimmer-30B-Q4_K_M.gguf` (~1.4 GB), Meta's official vision encoder
+ projector for Muse Glimmer 30B.

### Usage (llama.cpp)

```bash
huggingface-cli download mlasli/Muse-Glimmer-30B-Heretic-Abliterated-Q6_K-GGUF \
  --include "Muse-Glimmer-30B-Heretic-Abliterated-Q6_K.gguf" \
  --include "mmproj-Muse-Glimmer-30B-Q4_K_M.gguf" \
  --local-dir ./models

./build/bin/llama-mtmd-cli \
  -m ./models/Muse-Glimmer-30B-Heretic-Abliterated-Q6_K.gguf \
  --mmproj ./models/mmproj-Muse-Glimmer-30B-Q4_K_M.gguf \
  --image photo.png \
  -p "Describe this image."
```

> **Ollama note**: Ollama does not currently support separate `mmproj` files
> for this architecture. For image input, use llama.cpp (`llama-mtmd-cli` or
> `llama-server --mmproj`).

## License

Apache 2.0 (same as base model)

## Changelog

### v1.1.0 — vision (multimodal) support (2026-08-16)

- Added `mmproj-Muse-Glimmer-30B-Q4_K_M.gguf` (~1.4 GB), Meta's official vision encoder + projector,
  enabling image input via llama.cpp.
- The vision tower is untouched by abliteration, so this projector matches the
  base model (`meta-models/Muse-Glimmer-30B`).
- v1.0.0 was the initial (unversioned) text-only upload.