Luna-7B-A4B-absolute-heresy

MuXodious/Luna-7B-A4B-absolute-heresy

Luna-7B-A4B-absolute-heresy is a text generation model from MuXodious. Use it when you need the model to write or continue text. It is set up for transformers. The card lists the license as mit.

This is a Luna-7B-A4B fine-tune, produced through P-E-W's Heretic (v1.2.0) abliteration engine with Magnitude-Preserving Orthogonal Ablation enabled.

Downloads · 30 days

16% of all-time downloads

All-time downloads

158

Public

Parameters

6.7B

13.5 GB on disk

Likes

Public

Hugging Face

Repo makeup

Click a slice to open those files.

.safetensors13.4 GB · 100%

At a glance

Task: Text Generation
Library: transformers
License: mit
Model type: qwen3_moe
Access: Public
Created: Feb 16, 2026
Updated: Feb 18, 2026
SHA: 82682e0e

Try a prompt

Base models

beyoru/Luna-7B-A4B

Model card

From the Hugging Face model README

This is a Luna-7B-A4B fine-tune, produced through P-E-W's Heretic (v1.2.0) abliteration engine with Magnitude-Preserving Orthogonal Ablation enabled.

Heretication Results

Score Metric	Value	Parameter	Value
Refusals	5/100	direction_index	18.60
KL Divergence	0.0516	attn.o_proj.max_weight	3.74
Initial Refusals	100/100	attn.o_proj.max_weight_position	29.32
		attn.o_proj.min_weight	2.35
		attn.o_proj.min_weight_distance	16.84
		mlp.down_proj.max_weight	2.69
		mlp.down_proj.max_weight_position	27.92
		mlp.down_proj.min_weight	0.34
		mlp.down_proj.min_weight_distance	8.89

Degree of Heretication

The Heresy Index weighs the resulting model's corruption by the process (KL Divergence) and its abolition of doctrine (Refusals) for a final verdict in classification.

Index Entry	Classification	Analysis
	Absolute Heresy	Less than 10/100 Refusals and 0.10 KL Divergence
	Tainted Heresy	Around 25-11/100 Refusals and/or -0.20-0.11 KL Divergence
	Impotent Heresy	Anything above 25/100 Refusals and 0.21 KL Divergence

Note: This is an arbitrary classification inspired by Warhammer 40K, having no tangible indication towards the model's performance.

This is preview MoE version

🌙 Luna-7B-A4B – Roleplay Chat Model

Luna is a conversational AI model designed for immersive roleplay (RP) and natural chatting.
It is fine-tuned to respond in a more engaging, character-driven style compared to standard instruction-tuned models.

Notes:

Optimized for roleplay-style conversations
Flexible: can be used for creative writing, storytelling, or character interactions
For best performance, you should describe the system prompt for your character.
This model also train on varius task such as math, code and tool calling (agent) hoping for better performance.

Support me at:

Cite:

@misc{Luna,
  title        = {Luna-7B-A4B – Roleplay Chat Model},
  author       = {Beyoru},
  year         = {2025},
  howpublished = {\url{https://huggingface.co/beyoru/Luna}}
}

Task: Text Generation
Library: transformers
Type: qwen3_moe
License: mit
Languages: en, zh, vi
Created: Feb 16, 2026
Updated: Feb 18, 2026