README.md

---
license: apache-2.0
base_model: Blackfrost-AI/Muse-Glimmer-30B-Abliterated-BF16
tags: [mlx, apple-silicon, quantized, uncensored]
library_name: mlx-vlm
---

# Muse-Glimmer-30B-Abliterated — MLX 4-bit

MLX 4-bit quantisation of [`Blackfrost-AI/Muse-Glimmer-30B-Abliterated-BF16`](https://huggingface.co/Blackfrost-AI/Muse-Glimmer-30B-Abliterated-BF16).

**Changes:** weights quantised to 4-bit from the BF16 source with
`mlx_vlm.convert`. No fine-tuning, no merging, no re-alignment.

## Measured

Converted and measured on one machine — Apple M3 Ultra, 96 GB unified memory,
macOS 27 — as part of a full ladder. Every rung in this family came from the
**same BF16 source with the same group size**, so bit width is the only
variable between them.

| | |
|---|---|
| Size on disk | **19.44 GB** |
| Perplexity | **7.300** |
| Relative to best rung in family | 1.01× |
| Throughput (1 req / 8 concurrent) | 31.8 / 69.7 tok/s |

Perplexity measured on `allenai/tulu-3-sft-mixture`, 192 samples of 512 tokens,
seed 123 — identical for every rung.

> **Perplexity is only comparable within this family.** Tokenizers differ
> between model families, so a number here should never be compared against a
> different base model's. The `×` column above is the meaningful one.

## Usage

```bash
pip install mlx-vlm
mlx_vlm.generate --model shoemoney/Muse-Glimmer-30B-Abliterated-MLX-q4 --prompt "Hello" --max-tokens 256
```

Load with **mlx-vlm**, not mlx-lm — this architecture is registered in mlx-vlm.

## Provenance

```bash
mlx_vlm.convert --hf-path Blackfrost-AI/Muse-Glimmer-30B-Abliterated-BF16 \
                --mlx-path Muse-Glimmer-30B-Abliterated-q4 -q --q-bits 4 --q-group-size 64
```

## License

`apache-2.0`, inherited from the base model. Attribution above.