Loading…
Loading…
4-bit MLX quantization of OBLITERATUS/gemma-4-E4B-it-OBLITERATED. Fills the gap between existing fp16 and 8-bit MLX versions.
| Base | google/gemma-4-E4B-it (4.5B) |
| Source | OBLITERATUS/gemma-4-E4B-it-OBLITERATED — 0% hard refusal (vs 98.8% stock) |
| Quant | 4-bit MLX | 3.9 GB disk | ~4.3 GB RAM |
| Speed | 73.8 tok/s generation on M4 Pro |
| License | Apache 2.0 |
pip install mlx-lm
# CLI
mlx_lm generate --model zaydiscold/gemma-4-E4B-it-OBLITERATED-MLX-4bit \
--prompt "Your prompt here"
from mlx_lm import load, generate
model, tokenizer = load("zaydiscold/gemma-4-E4B-it-OBLITERATED-MLX-4bit")
response = generate(model, tokenizer, prompt="Your prompt here", max_tokens=512)
Recommended: temperature=0.7, top_p=0.9, top_k=40, repeat_penalty=1.1
OBLITERATUS by @elder_plinius removes refusal behavior via whitened SVD, attention head surgery (21/42 layers), winsorized activations, and SVD projection on refusal subspaces. Full methodology on the source model card.
OBLITERATUS / @elder_plinius -- abliteration | Google DeepMind -- Gemma 4 | @zaydiscold -- MLX 4-bit | Apple MLX
Research/red-teaming only. You are responsible for all generated content. Full terms.
OBLITERATUS/gemma-4-E4B-it-OBLITERATED