← back to catalog · registered 2026-09-23 19:57

TheDrewBenson/gemma-4-26B-A4B-it-uncensored-heretic-ara-MLX-8bit-int8-affine-betterCache

TheDrewBenson Gemma 26B MoE multimodal second-order
curl -H "Authorization: Bearer $ABL_KEY" \
     "https://abliteration.org/api/v1/models/TheDrewBenson%2Fgemma-4-26B-A4B-it-uncensored-heretic-ara-MLX-8bit-int8-affine-betterCache"
Response includes
  • classification m3
  • files 15
  • author_summary 1 models
  • readme_text full
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
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)
Refusal direction extracted via
Extraction technique

Difference-of-means

Confidence
HIGH
Why we say so
name contains 'heretic'; Heretic default extraction is difference-of-means (Arditi 2024)
Downloads · 30-day
0
Likes
0
Model age
today
created 2026-09-23

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.

Metadata

License
apache-2.0
Tags
mlx safetensors gemma4 heretic uncensored decensored abliterated ara image-text-to-text conversational en base_model:zecanard/gemma-4-26B-A4B-it-uncensored-heretic-ara-MLX-8bit-int8-affine
Total size
27.5 GB
Files
15
Quantizations
1
Registered
2026-09-23 19:57
Last updated on HF
2026-09-23 19:13

Files by quantization

Auxiliary files 15 files 27.5 GB
model-00005-of-00006.safetensors 4.87 GB b4f86fb0 download
model-00003-of-00006.safetensors 4.87 GB 573d1e3a download
model-00002-of-00006.safetensors 4.87 GB 830cb9ed download
model-00004-of-00006.safetensors 4.81 GB 222625ef download
model-00001-of-00006.safetensors 4.77 GB 5f8199c9 download
model-00006-of-00006.safetensors 3.32 GB 0ec1bc44 download
tokenizer.json 30.7 MB a2619fe1 download
model.safetensors.index.json 173 KB e6453409 download
chat_template.jinja 18.0 KB e46da626 download
config.json 12.0 KB 5707179e download
tokenizer_config.json 2.75 KB dc684e39 download
README.md 2.69 KB 5a4d5721 download
.gitattributes 1.53 KB 52373fe2 download
processor_config.json 902 B 13e92a44 download
generation_config.json 203 B edda3c19 download

README current version from Hugging Face


base_model: zecanard/gemma-4-26B-A4B-it-uncensored-heretic-ara-MLX-8bit-int8-affine
language: en
library_name: mlx
license: apache-2.0
license_link: https://ai.google.dev/gemma/docs/gemma_4_license
pipeline_tag: image-text-to-text
tags:

  • mlx
  • heretic
  • uncensored
  • decensored
  • abliterated
  • ara

🦆 TheDrewBenson/gemma-4-26B-A4B-it-uncensored-heretic-ara-MLX-8bit-int8-affine-betterCache

This model builds on zecanard/gemma-4-26B-A4B-it-uncensored-heretic-ara-MLX-8bit-int8-affine, which was converted to MLX from trohrbaugh/gemma-4-26B-A4B-it-heretic-ara using mlx-vlm version 0.6.3.
Please refer to zecanard's model card or the original model card for more details.

🌟 Quality

Quantized vision language model with an effective 9.163 bits per weight.

mlx_vlm.convert --quantize --q-group-size 32 --q-bits 8 --q-mode affine

🛠️ Customizations

This quant includes a bugfix for tools calling. It is aware of the current date, and also enables thinking (if available). You may disable this behavior by deleting the following line from the chat template, or changing true to false:

{%- set enable_thinking = true %}

You may need to adjust your environment’s Reasoning Section Parsing to recognize <|channel>thought as the Start String, and <channel|> as the End String.

Prompt-caching fix

We built on top of zecanard's work, removing the current time of day from the
chat template's date line to allow for prompt caching. The original template
rendered The current date is: ... down to the second (plus a raw Unix
epoch value) as the very first thing in every prompt, so it changed on every
single request and prevented any prefix/KV cache reuse between turns — every
request paid full prompt-processing cost no matter how much of the
conversation was unchanged. This fork renders date-only, so the line stays
stable across requests made on the same day and cache hits work normally
again, with no observable change in model behavior.

🖥️ Use with mlx

pip install -U mlx-vlm
mlx_vlm.generate --model TheDrewBenson/gemma-4-26B-A4B-it-uncensored-heretic-ara-MLX-8bit-int8-affine-betterCache --max-tokens 100 --temperature 0 --prompt "Describe this image." --image <path_to_image>
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 app" button that hands off directly to a local runtime of your choice - Abliteration, LM Studio, or Ollama. No API keys, no subscription, no prompt leakage.