library_name: mlx
pipeline_tag: text-generation
base_model: King3Djbl/mythos-9b-merged
tags:
- mlx
- apple-silicon
- abliterated
- uncensored
- heretic
- mythos
🍎 Mythos-9B-Merged · Abliterated · MLX
Uncensored King3Djbl/mythos-9b-merged in MLX 4-bit for Apple Silicon. A roomy Claude Mythos merge, fully local and private, no NVIDIA, no cloud, no API. Built for Mac developers and creators who want an uncensored local model that runs fast on their MacBook.
🚀 Run on your Mac
LM Studio: search richardyoung/mythos-9b-merged-abliterated-mlx and load it.
mlx-lm (terminal):
pip install mlx-lm
mlx_lm.chat --model richardyoung/mythos-9b-merged-abliterated-mlx
💾 Hardware
MLX 4-bit, ~5 GB. Apple Silicon (M1-M4), no NVIDIA/CUDA needed.
| Mac RAM | Fit |
|---|---|
| 8 GB | tight but runs |
| 16 GB+ | comfortable, room for long context |
Even a base 16 GB MacBook Air runs this 9B at usable speed.
📊 Abliteration
| Metric | Value |
|---|---|
| Base | King3Djbl/mythos-9b-merged (Qwen3, 9B) |
| Refusals after | 7 / 100 |
| KL divergence | 0.021 (minimal collateral damage) |
| Method | Heretic directional ablation |
| Format | MLX 4-bit |
🧬 What is abliteration?
Heretic removes the model's internal refusal direction from the weights, so it stops declining requests, with no retraining and capabilities intact.
⚠️ Disclaimer
Safety guardrails removed. For research and authorized use only.
🙏 Acknowledgments
Heretic by Philipp Emanuel Weidmann · MLX by Apple · Base by King3Djbl
Built & maintained by Richard Young · DeepNeuro