---
language:
  - en
  - ru
license: apache-2.0
library_name: mlx
base_model:
  - WaveCut/Qwythos-9B-v2-Heretic
base_model_relation: quantized
tags:
  - heretic
  - uncensored
  - abliteration
  - mlx
  - apple-silicon
  - qwen3.5
pipeline_tag: text-generation
---

# Qwythos-9B-v2-Heretic-MLX-8bit

**MLX 8-bit quantization** of [`WaveCut/Qwythos-9B-v2-Heretic`](https://huggingface.co/WaveCut/Qwythos-9B-v2-Heretic) — the Heretic-decensored version of [`empero-ai/Qwythos-9B-v2`](https://huggingface.co/empero-ai/Qwythos-9B-v2). Built for **Apple Silicon** (M1/M2/M3/M4).

## Specs

| Field | Value |
|---|---|
| Bits/weight | **8.501** |
| File size | ~9.5 GB |
| Minimum RAM | ~11 GB unified memory |

## Quantization

| Step | Tool | Version |
|---|---|---|
| Convert + quantize | `mlx_lm.convert` | **mlx-lm 0.31.3** (mlx 0.31.x) |
| Quant mode | `affine` (default) | `-q --q-bits 8` |

```bash
python -m mlx_lm.convert \
  --hf-path WaveCut/Qwythos-9B-v2-Heretic \
  -q --q-bits 8 \
  --upload-repo WaveCut/Qwythos-9B-v2-Heretic-MLX-8bit
```

## Usage

```python
from mlx_lm import load, generate
model, tokenizer = load("WaveCut/Qwythos-9B-v2-Heretic-MLX-8bit")
response = generate(model, tokenizer, prompt="Hello", max_tokens=256)
print(response)
```

```bash
# CLI
mlx_lm.generate --model WaveCut/Qwythos-9B-v2-Heretic-MLX-8bit --prompt "Hello"
```

## Architecture

Qwen3.5 **hybrid** — 32 blocks mixing attention and SSM (Mamba-style) layers. Supported in mlx-lm ≥ 0.31.0.

## Disclaimer

Uncensored (safety alignment removed via Heretic). The original `empero-ai/Qwythos-9B-v2` maintainers are not affiliated with this derivative. Use responsibly.