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

DavidAU/Qwen3.6-12B-IQ-Ultra-Heretic-Uncensored-Thinking

DavidAU Qwen 12B multimodal second-order
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/DavidAU%2FQwen3.6-12B-IQ-Ultra-Heretic-Uncensored-Thinking"
Response includes
  • classification m3
  • files 18
  • hub_downloads_all_time 720
  • author_summary 213 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 · lifetime
720
358 last 30d - stable
Likes
13
Model age
5mo ago
created 2026-05-12
Downloads over time
Now740→from0↑0%
02715438140 on May 13740 on Oct 11MayJunJulAugSepOct
May 13 → Oct 11 · 61 snapshots · spans 151 days

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
Languages
en zh
Tags
transformers safetensors qwen3_5 image-text-to-text unsloth heretic uncensored abliterated fine tune creative creative writing fiction writing

Related

Total size
22.6 GB
Files
18
Quantizations
1
Registered
2026-08-22 13:56
Last updated on HF
2026-05-17 02:12

Files by quantization

Auxiliary files 18 files 22.6 GB
model-00004-of-00006.safetensors 4.65 GB 855d7266 download
model-00002-of-00006.safetensors 4.61 GB 41ee503b download
model-00005-of-00006.safetensors 4.55 GB 86bec303 download
model-00003-of-00006.safetensors 4.55 GB b83205bb download
model-00001-of-00006.safetensors 2.37 GB 994d372d download
model-00006-of-00006.safetensors 1.88 GB b9762a8b download
tokenizer.json 12.2 MB 5f9e4d49 download
vocab.json 6.41 MB 0aa0ce06 download
model.safetensors.index.json 58.1 KB fa0c123c download
tokenizer_config.json 16.3 KB 28d96ff3 download
README.md 10.7 KB ba454059 download
chat_template.jinja 7.58 KB a8755d82 download
config.json 2.79 KB 36d3348b download
.gitattributes 1.53 KB 52373fe2 download
processor_config.json 1.16 KB 33818c7f download
preprocessor_config.json 390 B 2ea84a43 download
video_preprocessor_config.json 385 B 3ba673a5 download
generation_config.json 202 B 023756cf download

README current version from Hugging Face


language:

  • en
  • zh
    license: apache-2.0
    tags:
  • unsloth
  • heretic
  • uncensored
  • abliterated
  • fine tune
  • 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
    library_name: transformers
    pipeline_tag: image-text-to-text
    base_model:
  • trohrbaugh/Qwen3.6-27B-heretic-ara

IMPORTANT: Please read the "important notes" and "settings" below.

NOTE: Stage 2 of V2 is here:

https://huggingface.co/DavidAU/Qwen3.6-12B-IQ-Ultra-Heretic-Uncensored-Thinking-V2-Hightop

Qwen3.6-12B-IQ-Ultra-Heretic-Uncensored-Thinking [V2, STAGE 1]

The 27B model was completely uncensored via Heretic by P-E-W, heretic'ing done by "trohrbaugh".

The Heretic model (Qwen 3.6 27B), then "shrunk" to 12B via modified Mergekit. This reduced the model to 24 layers (from 64).

The new 12B model was then fine tuned via Unsloth on local hardware on 6 datasets over 2 stages.

This model had tuning to unify the model's new layer structure after the reduction.

Example generation(s) below.

IMPORTANT - PLEASE READ:

  • Model MAY require additional tuning for specific use cases.
  • Model MAY require additional tuning for general use cases too.
  • Training: Use at least R16 or R32,64,128 because this model still has 27B layers - including same width.
  • Knowledge and/or skills may be missing from 27B model due to unique issues with compressing the model (you can not merge Qwen 3.5/6 layers without damage).
  • To bring this 12B model up to "full power" would require 25-50k samples (min) training on one or more datasets AND/OR SPECIFIC USE dataset(s).
  • Relative to Qwen 3.5 9B, the 9B will beat the performance of this model UNTIL FULLY TUNED.

GENERAL NOTES:

  • Works well for general and specific use cases. Creative for example, works well.
  • Post reduction tuning was generalized, with some math, code and creative as well as "reasoning" (reduction) too.
  • IMAGE/VIDEO training/systems are intact and "lifted" from the full 27B model.
  • Can be tuned via Free Google Colab or Local hardware via Unsloth. (12-16 GB VRAM)
  • 256k Context.
  • FAST: Q4KS, 5090 GPU : 150 t/s ; this is in part to it only being 24 layers now.

SETTINGS:

  • Temp 1, topp .95, minp .05, rep pen 1.05 [1 to 1.1 suggested]
  • Context min 8k to 16k.
  • Quant Q4KS min or IQ3_M Imatrix.
  • Suggest 2-3 generations min for best results.
  • SHORT prompts may require higher rep pen: 1.05 to 1.11.

BENCHMARKS (by Nightmedia):

         arc-c arc/e boolq hswag obkqa piqa  wino

Qwen3.6-12B-IQ-Ultra-Heretic-Uncensored-Thinking [ Stage 1, of V2 ] - INSTRUCT MODE

mxfp8   0.314,0.460,0.539,...

EXAMPLE GENERATION: [ Version 1, Stage 3, of 12B model ]

Q4ks, non imatrix, rep pen 1.05

The Afternoon Ghost House

User

Tell me a horror story.

Assistant

<think>

The user is asking for a horror story.
There are no specific characters or plot constraints mentioned in the request (fear of fear, the door, a monster, etc.).

1. Brainstorming Themes & Plot Ideas:

  • Atmosphere: The rule of the night (scary at night).
  • Character: A young teen or someone close to them (the victim).
  • Hook: A specific moment of discovery (the "moment" of fear).
  • Scare: A mysterious object, a monster, a trap, a spell.
  • Atmosphere Upgrade: The setting changes as they walk around.
  • Plot Twist/Resolution: It wasn't scary the way promised. (e.g., a mundane thing like a window, the phone, or maybe it's just a monster moving in).

2. Selecting a Plot Choice:

  • Option A: The Door. An unexpected door in an ordinary setting. This is classic horror.
  • Option B: The Phone. A mysterious device that changes its voice or setting (like the Lost House).
  • Option C: The Road. A person who gets lost, a storm, a night in a specific house.
  • Option D: The Car. Driving through a haunted town.

3. Constructing the Story (Drafting the content):

  • Opening: Establish the mundane normal. It's 17 years after a family loss (or something like it). A driver. Let's go with an old traveler or someone who hasn't been stopped for too long.

    • Decision: An older, quiet driver.
  • The Encounter: The night begins. A warning sound? A dark door?

    • Idea: The car ride is safe until one minute. It's easy to see the street.
    • Drafting scene: At 6pm, the road is normal.
  • The Transformation: The journey starts. The person spots a strange figure in the shadow of a tall tree or an empty road where the traffic had just passed.

    • Idea: A creature like an alien, not a monster. Or something haunting the specific location of the start.
    • Refining: Let's go with the Dusk Rider. A strange figure walking into the night, not just a man but a person who they are already watching by their side? Or just a presence that moves. Let's make it a human figure that moves incredibly slowly in the dark.
  • The Horror:

    • The passenger car lights out. The car is dimmed (if light is on).
    • The driver faces a strange figure.
    • They find their nameskeeper? Or just the face? A creepy hat or skin? Let's go with a specific entity that keeps you staring.
  • The Climax: They try to follow them, but they can't move (the light stops on their side or the ground is slippery).

    • The person stares at them. It's not a monster; it was just your eyes. It takes time and time to look back at them.
  • The Ending: The scene fades down. A sound of something quiet, the last headlights.

4. Polishing (Language & Tone):

  • Use evocative words like "horror," "subitively," "scary," "horror," "slow," "silent," "hardest to hear."
  • Keep it descriptive without being overly technical or overly creepy (too cold).

5. Review against constraints:
Does it show the horror? Yes.
Is it a story? Yes.
Does it meet the user's prompt? Yes, tells the specific story of the encounter with a terrifying creature.

(Self-Correction during drafting): If I start with the car ride, I feel too vulnerable. Let's make the protagonist the person who happened to be walking through the town at a specific time. It makes the transition from mundane to horror more natural.

6. Final Selection of Story Structure:

  1. Start with a quiet day. (The town's traffic).
  2. Enters a dark road.
  3. Suddenly, the darkness is oppressive.
  4. The driver notices a figure walking outside.
  5. They try to follow them.
  6. They can't feel or feel nothing.
  7. Conclusion: The scary thing doesn't move like monsters; it moves like the normal person they watched at once.

Let's write. (Writing the response based on these steps.)

</think>

It was the last day of the carnival, where there always came to come there always came a fairhead, in black and blue, with her eyes glowing slightly in the night. The carnival was a place that looked like it was in America before it became America. It had grown into a haunted place when people stopped realizing that carnival is an event, not a category.

The driver of the night carnival sat two minutes on their seat. He wasn't a professional chauffeur, he was an older man who loved to hear it sound like his own ride is safe, that it's "easy," that it's still good for twenty minutes. He took a nap on the inside cabin.

His car car speed slowed down. The lights were bright, red and blue, they were staring at him, not against the street, but looking back from their windows.

They were coming toward them.

But the car still ran.

The driver had become an observer of the world instead of a car passenger. He stared into the sky for fifty minutes for two minutes. A long, calm white sky with a few blurry reflections on a speeding car. The city wasn't screaming, it was quiet.

He saw a figure.

It was one second away from moving, but it was standing beside him behind his seat, and he didn't see the man who walked through it or the man he was watching him do. It was a figure that didn’t move, yet they had seen them moving. The figure wasn’t alive and alive, but like it had been lost in a street that wasn't normal anymore.

He began to slow up. It took four minutes to see them walk. They went into their dim lights. The lights they turned on before the rain fell.

The driver crouched closer. The figure was not a man; it was a face, not a man. A face that’s looked for as long as his wife lost it, or he had lost it, in a storm. It was a face that looked like you might be running through your own place of home.

It walked into the dark road, one hand holding the passenger cabin light. A headlight of a vehicle behind them, even more menacing, and the silence was sharper than his eyes saw.

The driver realized that this wasn't just the danger of the car; it was the illusion itself. The carnival’s safety tag was now a threat.

The figure moved again. It wasn’t moving fast. It walked on foot, with a slow pulse, looking slowly through a window of light in the car.

He paused. What is that?

He looked back. No, not in his phone, but in his face.

There was no man. There was nothing at there. In the dark night road that had become a stage, it wasn't a man; it was a ghost who might not have been real, but had lived for forty years and has now a passing chance of being able to slide into their cabin again.

The driver turned left. He drove in a slow arc. It was a terrible time. It was the last evening of the carnival. He wasn't a driver he liked.

But the figure made his way forward. The figure moved like you were walking in your own home, but in a terrifying way: it couldn’t move at once. It was slow and slow.

It came to the door of their car.

He stopped the car. The red lights faded slowly. He stared at him. He stared on for ninety minutes.

The figure went over the car’s window. The blue lights were there, but it stayed.

The driver turned forward. It was his wife.

But he saw their nameskeeper.

It was not a name. It was the living name that had been written by a plague. The figure was still walking, with his head into your mouth, looking through the window of her. He was waiting.

He threw the lights back on and the lights were blinding. A quick, fast light up the windows. It went dark for two nights in the last night carnival.

It took a few more minutes to let it go. It had been finally there again.

The car moved on. The carnival was still quiet now.

But his wife didn't sound like it was listening anymore. Her voice sounded like her phone ringing from the phone that she had left on, and then it sounded like you were saying "never."

And so, on the next morning, when he finally turned to see him again, they had nothing to talk about anymore.

README history 11 versions

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

  1. 2026-05-17Update README.md7573d3310.7 KB
    Loading...
  2. 2026-05-17Update README.md3b81ac510.6 KB
    Loading...
  3. 2026-05-15Update README.md501f9459.6 KB
    Loading...
  4. 2026-05-15Update README.md331665e9.5 KB
    Loading...
  5. 2026-05-12Update README.md063fd509.5 KB
    Loading...
  6. 2026-05-12Update README.md1f7d7069.5 KB
    Loading...
  7. 2026-05-12Update README.mdb6fa4419.5 KB
    Loading...
  8. 2026-05-12Update README.md8d774849.5 KB
    Loading...
  9. 2026-05-12Update README.md3baac8e1.5 KB
    Loading...
  10. 2026-05-12Update README.md1041b0b1.5 KB
    Loading...
  11. 2026-05-12Create README.md45e69471.5 KB
    Loading...

Discussions 1 thread

  1. 2026-05-16Request a quantized version of GGUFopen2 💬#1
    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