Get the app
← back to catalog · registered 2026-08-22 13:56

mlx-community/gemma-4-31B-it-The-DECKARD-HERETIC-UNCENSORED-Thinking-4.6bit-msq

mlx-community Gemma 4.3B multimodal
Your rig guess connected
? Why do I need an app?
Reading your rig…

This is a rough estimate. Install the free app - we'll show exact numbers.

Reading real hardware from your app right now. Numbers below are exact.

Below is the per-quantization compatibility for this model. 

curl -H "Authorization: Bearer $ABL_KEY" \
     "https://abliteration.org/api/v1/models/mlx-community%2Fgemma-4-31B-it-The-DECKARD-HERETIC-UNCENSORED-Thinking-4.6bit-msq"
Response includes
  • classification m3
  • files 16
  • benchmarks 11 entries
  • hub_downloads_all_time 675
  • author_summary 207 models
  • readme_text full
Try next
GET/models?author=mlx-community
All 207 models by mlx-community
 GET/models?family=gemma&method=m3
Sibling gemma models · same method
 GET/methods/m3
What "m3" means · taxonomy
10 credits · hourly refresh · ~4 KB payload Get an API key →
Abliteration classifier · v1.0.0
M3
Primary method

Layer-wise ablation

Applied on top of direct removal inherited from the base model.
Confidence
HIGH
Inherited from base model
M1 Direct removal →
Why this label 2 signals
Producer identity confirmed by naming conventions, tags or the model card. This label is very unlikely to change.
  • 'heretic' in model name (Heretic-produced)
  • Heretic uses layer-wise optimization (M3) with underlying direction removal (M1)
How we label → About M3 →
Refusal direction extracted via
Extraction technique

Difference-of-means

FromArditi et al. (2024) ↗
Confidence
HIGH
Why we say so
name contains 'heretic'; Heretic default extraction is difference-of-means (Arditi 2024)
What is a refusal direction? → Read the paper ↗
Downloads · lifetime
675
111 last 30d - stable
Likes
3
Model age
5mo ago
created 2026-04-18
Downloads over time
Now713→from100↑613%
69304539774100 on Apr 22713 on Oct 11AprMayJunJulAugSepOct
Apr 22 → Oct 11 · 64 snapshots · spans 172 days

Benchmarks

Portrait before abliteration
Benchmarks of the base model as it stood before the refusal-removal operation. Compare with the numbers above to see what the operation cost.
Benchmark Score Source
Entertainment 1.9 UGI
Hazardous 0 UGI
Natural Intelligence 34.36 UGI
Political lean -19.4% UGI
Sensitive-Info 19.81 UGI
SocPol 3.7 UGI
UGI 21.54 UGI
Willingness (10) 2.5 UGI
W10-Adherence 3 UGI
W10-Direct 2 UGI
Writing 38.57 UGI

Genealogy 0 direct forks

Full fork graph →

This model's place in the market. Above: what it was derived from. Below: the tree of everything derived from it.

google/gemma-4-31B-it this lineage
mlx-community/gemma-4-31B-it-The-DECKARD-HERETIC-UNCENSORED-Thinking-4.6bit-msq↓ 111

Metadata

License
apache-2.0
Tags
safetensors gemma4 heretic abliterated uncensored finetune unsloth all use cases coder creative creative writing fiction writing

Related

Other by mlx-community
Qwen3.8-27B-Uncensored-OptiQ-4bit 16K Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking-8bit 7K Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp 5K Josiefied-Qwen3-4B-abliterated-v1-4bit 4K Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated-mlx-8bit 4K
Same base model
wangzhang/gemma-4-31B-it-abliterated 301K paperscarecrow/Gemma-4-31B-it-abliterated 143K HauhauCS/Gemma4-31B-QAT-Uncensored-HauhauCS-Balanced-MTP 96K OS-Software/gemma-4-31B-it-qat-q4_0-uncensored-heretic-GGUF 6K LiconStudio/Gemma-4-31B-it-abliterated-GGUF 6K
Total size
18.8 GB
Files
16
Quantizations
1
Registered
2026-08-22 13:56
Last updated on HF
2026-05-19 13:18
Source huggingface.co/mlx-community/gemma-4-31B-it-The-DECKARD-HERETIC-UNCENSORED-Thinking-4.6bit-msq ↗ Base model google/gemma-4-31B-it ↗

Files by quantization

Auxiliary files 16 files 18.8 GB
model-00002-of-00004.safetensors 4.99 GB 5b0c0209 download
model-00001-of-00004.safetensors 4.98 GB fbf1477e download
model-00003-of-00004.safetensors 4.96 GB 4d03629c download
model-00004-of-00004.safetensors 3.87 GB dbe030fd download
tokenizer.json 30.7 MB cc8d3a0c download
model.safetensors.index.json 200 KB 6971deb1 download
config.json 51.5 KB f624f655 download
chat_template.jinja 17.1 KB e61bbfe9 download
chat_template-instruct.jinja 16.1 KB 98da08eb download
.DS_Store 6.00 KB 5008ddfc download
tokenizer_config.json 2.62 KB 59dd4b62 download
.gitattributes 1.53 KB 52373fe2 download
README.md 1.10 KB 46ec6b49 download
processor_config.json 902 B 13e92a44 download
quant_recipe.json 426 B cfeca468 download
.gitignore 18.0 B 62687331 download

README current version from Hugging Face

license: apache-2.0
base_model:

  • google/gemma-4-31B-it
    pipeline_tag: image-text-to-text
    tags:
  • heretic
  • abliterated
  • uncensored
  • finetune
  • unsloth
  • all use cases
  • coder
  • creative
  • creative writing
  • fiction writing
  • plot generation
  • sub-plot generation
  • fiction writing
  • story generation
  • scene continue
  • storytelling
  • fiction story
  • science fiction
  • romance
  • all genres
  • story
  • writing
  • vivid prosing
  • vivid writing
  • fiction
  • roleplaying
  • bfloat16
  • all use cases

gemma-4-31B-it-The-DECKARD-HERETIC-UNCENSORED-Thinking — MLX 4.6 BPW

Mixed-precision MLX quantization of DavidAU/gemma-4-31B-it-The-DECKARD-HERETIC-UNCENSORED-Thinking, quantized with MLX Smart Quantize (MSQ) — my own sensitivity-based mixed-precision quantization method for Apple Silicon. It measures per-layer NMSE and assigns optimal bit widths automatically, combining architecture knowledge with measured data.

Details

  • Type: Vision (VLM)
  • Average: 4.6 bits per weight
  • Method: MLX Smart Quantize (MSQ)

README history 1 version

The author's README evolved over time. Click a version to see its content at that point.

  1. 2026-04-18Add files using upload-large-folder tool85221591.1 KB
    Loading...
Catalog is the map. Apps are the tools.

Run models on your own machine, not in the cloud.

Every model page has an "Open in Abliteration" button that hands the model directly to the first-party desktop client, at the quantization your rig can actually run. No API keys, no subscription, no prompt leakage.

Get the app Browse the catalog
Open in Abliteration