8-bit MLX quantization of ChanLumerico/EXAONE-3.5-7.8B-Instruct-Yaho — a Korean gyaru-persona ('갸루귀신') fine-tune of LGAI-EXAONE/EXAONE-3.5-7.8B-Instruct (LoRA SFT + custom ORPO). Persona, behavior triggers, and the baked 갸루귀신 name anchor are verified to survive 8-bit quantization. Non-commercial (EXAONE AI Model License 1.1-NC).

The persona anchor is baked into chat_template.jinja — no system prompt needed.

pip install mlx-lm==0.29.1 "transformers==4.57.6"

python

from mlx_lm import load, generate
model, tok = load("ChanLumerico/EXAONE-3.5-7.8B-Instruct-Yaho-8bit")
p = tok.apply_chat_template([{"role": "user", "content": "오늘 시험 망쳤어…"}],
                            add_generation_prompt=True, tokenize=False)
print(generate(model, tok, prompt=p, max_tokens=512))

See the full model card / eval / interpretability and the project repo. Base: EXAONE-3.5 (LG AI Research).