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

cs2764/gemma-4-31b-it-6bit-abliterated-mlx

cs2764 Gemma 31B
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/cs2764%2Fgemma-4-31b-it-6bit-abliterated-mlx"
Response includes
  • classification m1
  • files 15
  • hub_downloads_all_time 4,367
  • author_summary 28 models
  • readme_text full
10 credits · hourly refresh · ~4 KB payload Get an API key →
Abliteration classifier · v1.0.0
M1
Primary method

Direct removal

No other method signals detected in this model.
Confidence
MEDIUM
Why this label 3 signals
Method inferred from partial signals - repository name, related files, or tag patterns. Producer identity not confirmed; label may sharpen or shift as we gather more evidence.
  • 'abliterated' in name/tags
  • is_gguf=0 (base model)
  • no specific method indicators - defaulting to M1 (most common)
Refusal direction extracted via
Extraction technique

Difference-of-means

Confidence
MEDIUM
Why we say so
primary_method=M1; difference-of-means is the reference extraction for M1/M3 (Arditi 2024)
Downloads · lifetime
4K
177 last 30d - cooling
Likes
1
Model age
6mo ago
created 2026-04-04
Downloads over time
Now4.4K→from2.3K↑89%
2.2K3K3.8K4.6K2.3K on Apr 154.4K on Oct 11AprMayJunJulAugSepOct
Apr 15 → Oct 11 · 65 snapshots · spans 179 days

Metadata

License
other
Tags
mlx safetensors gemma4 abliteration uncensored experimental license:other 6-bit region:us

Related

Total size
24.8 GB
Files
15
Quantizations
1
Registered
2026-08-22 13:56
Last updated on HF
2026-04-04 23:23

Files by quantization

Auxiliary files 15 files 24.8 GB
model-00003-of-00005.safetensors 5.10 GB 7d142cfd download
model-00004-of-00005.safetensors 5.08 GB 80bedb25 download
model-00002-of-00005.safetensors 5.03 GB d532e426 download
model-00001-of-00005.safetensors 5.01 GB 066bfc95 download
model-00005-of-00005.safetensors 4.57 GB 3920b712 download
tokenizer.json 30.7 MB cc8d3a0c download
model.safetensors.index.json 201 KB ba60778d download
chat_template.jinja 11.8 KB 33c51c2d download
config.json 5.51 KB cfbb25ce download
tokenizer_config.json 2.62 KB 59dd4b62 download
README.md 2.08 KB 694d0e05 download
.gitattributes 1.53 KB 52373fe2 download
abliteration_log.json 796 B 8902a407 download
processor_config.json 627 B 09dfe230 download
generation_config.json 208 B e605bb45 download

README current version from Hugging Face


tags:

  • mlx
  • abliteration
  • uncensored
  • experimental
    license: other

gemma-4-31b-it-6bit-abliterated

This model was created using the mlx-abliteration toolkit,
which is based on the FiditeNemini/mlx-abliteration project.

Base Model

Original model: ./models/gemma-4-31b-it-6bit (local path)

What is Abliteration?

Abliteration is a mechanistic interpretability technique that identifies and orthogonalizes the
"refusal direction" in a model's activation space, surgically removing refusal behavior without
full fine-tuning.

Abliteration Parameters

Parameter Value
Base model ./models/gemma-4-31b-it-6bit
Ablation method projection
Refusal vector policy per-layer
Refusal direction method projected
Ablation strength 2.0
Probed layers 5,11,17,23,29,35,41,47,53,59
PCA components (ablate-k) 1
Attention only True
Excluded modules none
Timestamp 2026-04-04T23:22:15 UTC

This model uses a Mixture-of-Experts (MoE) architecture. Abliteration targets only the attention projection weights (q/k/v/o_proj) to preserve expert routing quality.

⚠️ Disclaimer

This model is intended for research, experimentation, and testing purposes only.

  • This model may produce harmful, offensive, inappropriate, or otherwise objectionable content.
  • The abliteration process removes safety guardrails that were intentionally built into the original model.
  • Do not use this model in production systems, consumer-facing applications, or any context
    where harmful outputs could cause real-world harm.
  • The authors and contributors of this toolkit bear no responsibility for any misuse of this model
    or any harm caused by outputs generated by this model.
  • By using this model, you agree that you are solely responsible for ensuring its use complies
    with all applicable laws and ethical guidelines.

This model is shared purely for academic and technical exploration of model internals.

README history 1 version

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

  1. 2026-04-04Add files using upload-large-folder toolbcbb1922.1 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.

Open in Abliteration