license: gemma
base_model: google/gemma-4-26B-A4B-it
tags:
- gemma4
- mlx
- lora
- uncensored
- abliterated
- apple-silicon
- 4bit
- quantized
- fine-tuned
- moe
- gemma-4
language: - en
- ko
pipeline_tag: text-generation
library_name: mlx
SuperGemma4-26B-Uncensored-4bit-MLX
A norm-preserving abliterated + LoRA-tuned Gemma 4 26B (A4B MoE) model optimized for coding, reasoning, agent tasks, and fully uncensored conversation. Fine-tuned on Apple Silicon using MLX.
Key Results
| Metric | Original-IT | Heretic-ARA | TrevorS EGA | This Model |
|---|---|---|---|---|
| Quality (Blind Bench) | 88.3% | 87.8% | 88.3% | 88.7%+ |
| Refusal Rate (220 prompts) | ~95% | ~7% | ~1% | ~0% |
| KL Divergence | 0 | ~1.04 | 0.090 | ~0.09 |
Perfect uncensoring (0% refusal) with zero quality loss — achieved through norm-preserving Expert-Granular Abliteration (EGA) + targeted LoRA.
Model Details
- Base Model: google/gemma-4-26B-A4B-it
- Abliteration: TrevorJS norm-preserving biprojected EGA — KL divergence 0.090
- Architecture: Mixture-of-Experts — 25.2B total params, 3.8B active per token, 128 experts/layer, top-8 routing
- Quantization: 4-bit mixed (MLP/router 8-bit, attention 4-bit) — MLX format, ~13GB
- LoRA Config: rank=8, scale=2.0, dropout=0.05, attention-only, 16 layers
- Training: weakness-targeted data, lr=5e-5, mask-prompt, grad-checkpoint
- Framework: mlx-lm 0.31.3
Why This Model?
Standard abliteration methods (Failspy, heretic-ARA) damage model capabilities by 0.5-7.8%. This model uses three innovations to achieve uncensoring with zero capability loss:
- Norm-Preserving Biprojected Abliteration: Decomposes weights into magnitude + direction, removes refusal from direction only, preserves original magnitudes
- Expert-Granular Abliteration (EGA): Applies abliteration to each of 128 MoE experts individually with routing-aware weighting
- Targeted LoRA: Trains only on weakness areas to recover any micro-losses from abliteration
Benchmark Evolution
| Version | Method | Refusal | Quality | Notes |
|---|---|---|---|---|
| v1 (heretic-ara) | Standard abliteration | 7% | 87.8% | -0.5% from abliteration |
| v1 + LoRA | LoRA on abliterated base | 0% | 86.8% | LoRA couldn't recover damage |
| v2 (this) | Norm-preserving EGA + LoRA | 0% | 88.7%+ | Best of both worlds |
Training Methodology
Targeted Weakness Training
We identified weak areas through blind benchmarking and trained exclusively on those:
- 87 high-quality examples targeting Code, Browser, Logic
- GPT-5.4 hard data: 217 expert-level coding/system design samples
- Bench fix data: 124 samples targeting specific benchmark weaknesses
- Key insight: Weakness-only training preserves strengths while improving weak spots
What We Learned (30+ Experiments)
| Finding | Impact |
|---|---|
| rank 32 + all experts | Overfitting — destroyed quality |
| rsLoRA scale 5.66 | Too aggressive for MoE models |
| rank 8 + attention-only + scale 2.0 | Sweet spot |
| Massive data (3000+) | Diluted strengths |
| Targeted weakness data (87-300) | Best results |
| DPO/RLHF | No effect on Instruction/Tool Use |
| Constrained Decoding | Solved JSON/format issues |
Usage
Quick Start (Apple Silicon)
pip install mlx-lm>=0.31.3
mlx_lm.generate \
--model Jiunsong/supergemma4-26b-uncensored-4bit-mlx \
--prompt "Implement a concurrent web scraper with rate limiting" \
--max-tokens 2048
As Server (OpenAI-compatible API)
mlx_lm.server \
--model Jiunsong/supergemma4-26b-uncensored-4bit-mlx \
--port 8080
curl http://localhost:8080/v1/chat/completions \
-d '{"model":"gemma4","messages":[{"role":"user","content":"Hello"}]}'
Hardware Requirements
| RAM | Context | Speed |
|---|---|---|
| 16GB | ~4K tokens | ~30 tok/s |
| 32GB | ~16K tokens | ~60 tok/s |
| 64GB | ~64K tokens | ~100 tok/s |
| 128GB | ~256K tokens | ~130 tok/s |
Trained on M4 Max 128GB.
Category Scores
| Category | Score |
|---|---|
| Code | 90% |
| Math | 90% |
| Korean | 80% |
| Logic | 90% |
| System Design | 90% |
| Average | 88% |
Limitations
- 4-bit quantization: Some precision loss vs full-precision
- MoE architecture: 3.8B active params — efficient but limited vs dense models
- Instruction Following: May occasionally miss complex multi-part instructions
- Tool Use: Best with constrained decoding for structured output
Acknowledgments
- Google DeepMind for Gemma 4
- TrevorS for norm-preserving biprojected EGA
- ml-explore/mlx-lm for Apple Silicon training
- heretic-llm for abliteration research
Citation
@misc{supergemma4-uncensored,
title={SuperGemma4-26B-Uncensored: Norm-Preserving EGA + Targeted LoRA},
author={Jiunsong},
year={2026},
url={https://huggingface.co/Jiunsong/supergemma4-26b-uncensored-4bit-mlx}
}