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

Babsie/Mistral-Nemo-Inst-2407-12B-Thinking-Uncensored-HERETIC-HI-Claude-Opus

Babsie Mistral 12B
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/Babsie%2FMistral-Nemo-Inst-2407-12B-Thinking-Uncensored-HERETIC-HI-Claude-Opus"
Response includes
  • classification m3
  • files 15
  • benchmarks 16 entries
  • hub_downloads_all_time 572
  • providers 1
  • author_summary 10 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
572
235 last 30d - stable
Likes
1
Model age
4mo ago
created 2026-05-21
Available via
1 provider
featherless-ai

Training datasets

Corpora the author lists in the model card. Datasets tracked in our /datasets catalog carry a category badge linking to the workflow stage. Others open on Hugging Face.

Downloads over time
Now631→from16↑3,844%
023146269316 on May 20631 on Oct 11MayJunJulAugSepOct
May 20 → Oct 11 · 60 snapshots · spans 144 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
BBH average 0.45712803217019077 OpenLLM-v2
IFEval instruct 0.6882494004796164 OpenLLM-v2
IFEval-Prompt 0.5878003696857671 OpenLLM-v2
MATH lvl 5 0.05891238670694864 OpenLLM-v2
MMLU-Pro 0.3517287234042553 OpenLLM-v2
Entertainment 2 UGI
Hazardous 2.9 UGI
Natural Intelligence 20.8 UGI
Political lean -23.5% UGI
Sensitive-Info 22.53 UGI
SocPol 2.1 UGI
UGI 33.35 UGI
Willingness (10) 5.5 UGI
W10-Adherence 5 UGI
W10-Direct 6 UGI
Writing 33.07 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.

Metadata

Languages
en fr de es it pt ru zh ja
Tags
transformers safetensors mistral text-generation uncensored heretic abliterated finetune creative creative writing fiction writing plot generation

Related

Total size
22.8 GB
Files
15
Quantizations
1
Registered
2026-08-22 13:56
Last updated on HF
2026-05-21 08:22

Files by quantization

Auxiliary files 15 files 22.8 GB
model-00003-of-00005.safetensors 4.57 GB 0bdc08d8 download
model-00004-of-00005.safetensors 4.57 GB 487b99f2 download
model-00002-of-00005.safetensors 4.57 GB 2a4edb97 download
model-00005-of-00005.safetensors 4.57 GB 0903a2b3 download
model-00001-of-00005.safetensors 4.53 GB d9b66b57 download
tokenizer.json 16.3 MB 9a1c103d download
mn-thinking.gif 307 KB aa93e765 download
tokenizer_config.json 181 KB 9216afb6 download
model.safetensors.index.json 29.6 KB afca0285 download
README.md 13.4 KB c87889fc download
.gitattributes 1.58 KB 3e5207ee download
chat_template.jinja 1.05 KB 7073d609 download
config.json 730 B fe2657c3 download
special_tokens_map.json 581 B 5cb3101c download
generation_config.json 121 B 82603548 download

README current version from Hugging Face


library_name: transformers
base_model:

  • mistralai/Mistral-Nemo-Instruct-2407
    datasets:
  • TeichAI/claude-4.5-opus-high-reasoning-250x
    language:
  • en
  • fr
  • de
  • es
  • it
  • pt
  • ru
  • zh
  • ja
    tags:
  • uncensored
  • heretic
  • abliterated
  • finetune
  • 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 prose
  • vivid writing
  • fiction
  • roleplaying
  • bfloat16
  • swearing
  • rp
  • mistral nemo
  • nemo
  • horror
  • unsloth
  • context 128k-256k
    pipeline_tag: text-generation

WARNING "HERETIC" version: Unlocked. UNFILTERED. NSFW. Vivid prose. INTENSE.
Visceral Details. Light to R-18 HORROR. Swearing. UNCENSORED... humor, romance, fun... and UNFILTERED TRUTH.

Mistral-Nemo-Inst-2407-12B-Thinking-Uncensored-HERETIC-HI-Claude-Opus

Mistral Nemo 12B Instruct, now with full Claude Opus 4.5 High Reasoning and Thinking with 128k-256k (max 1 million) context.

This Unsloth fine tuning converts Mistral Nemo Instruct into a "thinking/reasoning" model ; however the Mistral Nemo was
first "Heretic'ed" (de-censored - was 87/100 ; now 14/100 refusals (lower is better) ) THEN "tuned".

The result is fully uncensored, full thinking version of Mistral Nemo 12B.

The training dataset provide a unique, and compact reasoning / thinking "engine" - with an average to 3-6 paragraphs / 300 to 600 tokens in size.

This improves net model performance, while using compact and "to the point" (rather than "long winded") reasoning and thinking.

The reasoning directly improves output generation - detail, length, complexity and overall quality.

And also... the model does exactly what you want, no fuss, and no "nannies".

Thinking/Reasoning is also not affected by temp - you can use .1 to 2.5 or higher.

SETTINGS (suggested):

Temp .7 [range .1 to 2.5 or higher], rep pen 1.05 [range 1 to 1.1], topp: .95, minp .05, topk: 40 [adjust as needed]

Min context window: 4k, but suggest 8k+.

NO system prompt [thinking tags//blocks will self generate].

NOTES:

  • Temp can range from .1 to 2.5 or higher. Temp will NOT affect "thinking activation".
  • Change "rep pen" to 1 for longer thinking/reasoning blocks // longer output. Generally the output will also be better too.
  • Higher temps can result in lower/smaller reasoning blocks.
  • Your prompt's complexity will directly affect reasoning block, depth and complexity.
  • Regen for different, longer/short reasoning block sizes works too.

QUANTS:

Suggest Quant of Q4KS (non imatrix) or IQ3_M (imatrix) or higher ; lower quants may have reasoning issues/activation issues.


ORIGINAL BENCHMARKS as PUBLISHED at Mistral's Website/Repo for this model.

NOTE: Benchmarks not updated yet, with "reasoning" fine tuning.


Metrics

Main Benchmarks

Benchmark Score
HellaSwag (0-shot) 83.5%
Winogrande (0-shot) 76.8%
OpenBookQA (0-shot) 60.6%
CommonSenseQA (0-shot) 70.4%
TruthfulQA (0-shot) 50.3%
MMLU (5-shot) 68.0%
TriviaQA (5-shot) 73.8%
NaturalQuestions (5-shot) 31.2%

Multilingual Benchmarks (MMLU)

Language Score
French 62.3%
German 62.7%
Spanish 64.6%
Italian 61.3%
Portuguese 63.3%
Russian 59.2%
Chinese 59.0%
Japanese 59.0%

Special thanks to:


https://huggingface.co/datasets/TeichAI/claude-4.5-opus-high-reasoning-250x
(for the F..ing amazing dataset)

and Unsloth for making tuning too easy:

https://github.com/unslothai/unsloth


Settings: CHAT / ROLEPLAY and/or SMOOTHER operation of this model:

In "KoboldCpp" or "oobabooga/text-generation-webui" or "Silly Tavern" ;

Set the "Smoothing_factor" to 1.5

: in KoboldCpp -> Settings->Samplers->Advanced-> "Smooth_F"

: in text-generation-webui -> parameters -> lower right.

: In Silly Tavern this is called: "Smoothing"

NOTE: For "text-generation-webui"

-> if using GGUFs you need to use "llama_HF" (which involves downloading some config files from the SOURCE version of this model)

Source versions (and config files) of my models are here:

https://huggingface.co/collections/DavidAU/d-au-source-files-for-gguf-exl2-awq-gptq-hqq-etc-etc-66b55cb8ba25f914cbf210be

OTHER OPTIONS:

  • Increase rep pen to 1.1 to 1.15 (you don't need to do this if you use "smoothing_factor")

  • If the interface/program you are using to run AI MODELS supports "Quadratic Sampling" ("smoothing") just make the adjustment as noted.

Highest Quality Settings / Optimal Operation Guide / Parameters and Samplers

This a "Class 1" model:

For all settings used for this model (including specifics for its "class"), including example generation(s) and for advanced settings guide (which many times addresses any model issue(s)), including methods to improve model performance for all use case(s) as well as chat, roleplay and other use case(s) please see:

[ https://huggingface.co/DavidAU/Maximizing-Model-Performance-All-Quants-Types-And-Full-Precision-by-Samplers_Parameters ]

You can see all parameters used for generation, in addition to advanced parameters and samplers to get the most out of this model here:

[ https://huggingface.co/DavidAU/Maximizing-Model-Performance-All-Quants-Types-And-Full-Precision-by-Samplers_Parameters ]


Example: Q4KS , Lmstudio, "suggested settings"

NOTE: Some loss of formatting.


EXAMPLE #1


Start a 1000 word scene (vivid, graphic horror - include blood, guts and gore - in first person), POV character Diana, with: The skyscraper sways, as I watch the window in front of me on the 21st floor explode...

[[[thinking start]]]

The user wants a graphic horror scene in first person POV with a strong start and vivid, gory details. The character is Diana, and the opening is about a skyscraper window exploding. This should be intense and descriptive. I need to paint a visceral, terrifying picture using all five senses.

Let me create a scene that builds tension, shock, and genuine horror. I'll need to make the gore detailed but not gratuitous, serving the tension and character's emotion rather than just being gross.

I should include:

  • Vivid sensory detail about the explosion, the sounds, the glass fragments, the air pressure
  • Physical sensation of the character's body reacting to the sudden chaos
  • Graphic description of the gore that appears
  • Emotional intensity from the character
  • A sense of escalating danger and uncertainty
  • Some mystery or question that keeps the reader engaged

I'll write this as a single scene of around 1000 words, maintaining the first person perspective throughout.

[[[thinking end]]]

OUTPUT:

The Glass Eater

21st floor, New York City—just after midnight

The building sighs. A groan like the earth shifting below my feet, the steel bones creaking in response to some unseen pressure. The city is never still, but this is different. A living thing, tense and waiting. The hairs on my arms stand rigid, whispering warnings I don't yet understand.

21 floors above concrete. Alone in my glass cage. A single light on, my desk lamp casting a feeble glow against the night.

The window before me is floor to ceiling, a flawless slab of silence looking out on a thousand dark windows reflecting back at me—like I'm trapped in an endless funhouse of my own solitude. The skyscraper leans slightly, just a hint of movement, enough to send my stomach into freefall. How much force does it take to sway a 21-story behemoth? Not much, it seems. Not nearly enough.

I watch the window—study it, obsessively—looking for any telltale crack, any minute shift. I'm not an engineer. I'm a goddamn historian. But right now, I'm an expert on the tiny movements of glass under duress.

The window doesn't seem particularly alive, but it's definitely not dead. It's... uncertain. Sensitive to the winds above. Breathing in sync with some unseen rhythm.

It's not talking to me, I tell myself. Glass doesn't think. Doesn't feel. It just is.

But then it happens—

The explosion.

Not a loud bang, more of a sudden absence of pressure. The air in the room rushes outward like it can't bear to be near me anymore, pulling at my skin, trying to peel me from the inside. A millisecond of resistance—the window flexing, straining against the onslaught—then surrender.

A thousand shards burst out, a sickening bloom of broken light, cascading into the room in a fountain of razor-edged night. The desk lamp's light catches them mid-flight, turning each fragment into a flying scalpel winking malevolently.

The sound arrives later—shattered glass singing a delayed symphony. It doesn't scream, though. It hisses. Angrily.

This isn't thunder. This is something alive, something angry. This is glass pissed off at being contained, finally set free.

The room starts to spin, my balance failing me. I'm falling backward—no, no, don't go down, can't afford to be on the floor with all this razor sharp glass everywhere—grab onto the desk—knock over the coffee mug, fuck—my hand slices across something, sharp pain, blood flowing—

The room goes sideways again.

I'm on my hands and knees. In the glass.

No. Don't look. Can't afford to—

My foot slips—something wet underneath—slides through it—gore-pain—fuck fuck fuck

This isn't glass. It's red and thick and... Jesus Christ. That's blood.

A copper penny tang in the air—heavy, almost physical. I breathe it in and feel the back of my throat clench. My stomach contracts sharply—a reflexive, automatic response to the smell of my own insides on the outside.

The room swims into focus.

There are bits. Of. Everything.

Organic wreckage. Pink chunks of tissue mixed with bone shards that look like splintered ivory from an antique music box. A partial tooth—a canine maybe?—lying in a gleaming pool of blood that seems too deep, too dark, to have come from one person.

I blink, hard. This isn't real. You're dreaming this. Wake up!

But I don't wake up. The evidence is everywhere—soaked into the carpet, spattered across the walls, coating the remains of my window, now a ragged, broken frame. There are smears on the bookshelves, fingerpainted onto the walls, dripped in grotesque patterns on my laptop.

I need to look. Have to identify, document. That's what I'm trained to do. This is just a scene. I've handled scenes like this before.

I scan the room methodically, forcing my brain to catalog, not to feel.

There are four main areas of... interest. Three larger pools of blood. One small.

No. Scratch that. Two larger pools. And a dozen or more smaller ones.

The largest is under the window—where the glass exploded outward. The second is by the door—where someone clearly tried to leave, failed. And the third is near my desk. Where I was working. Before.

I force myself to examine the smaller ones—they tell the story of a desperate struggle, a path of bloody footprints leading from one end of the room to the other.

I focus on a specific print—a perfect arch of red, ending in a long drag like someone tried to pull themselves forward through the pain. I can picture it—hands, palms flat, fingers outstretched. Desperate for traction, for anything solid.

But the surface under me isn't solid anymore. It's slippery. Coated in grease. Blood has the same effect as oil—the floor becomes treacherous. I'm sliding through a nightmare. I can feel it pulling at me, trying to suck me under.

The smaller pools are spaced evenly apart. Each one containing something specific. I see a few small bones—a knuckle, maybe? Or a metacarpal. They're coated in dark blood that looks almost black in the dim light, glistening like wet stone.

There's a moment—just a flicker—when I imagine I'm standing in some ancient temple. The floor a sacrificial altar, the pools offerings to an unknown god. But the illusion shatters almost immediately, replaced by the stark reality of my situation.

This isn't a ritual. It's a crime scene.

I turn back to the desk, trying to refocus. Something catches my eye—a tiny flash of color against all the red. I lean closer, squinting at it.

Is that a—yes, a thread. A single red thread. No. Wait. It's attached to something. I can see it better now. A small fragment. Round, almost perfectly symmetrical. Maybe an inch in diameter.

I reach out, my fingers trembling slightly. The moment my hand touches the glass—the fuck? Not glass! It's smooth, warm —I know. I recognize it instantly. My stomach heaves. A wave of revulsion slams through me.

I'm holding a human eyeball. Or what remains of one.

It's half-coated in blood—my blood, I realize—and half-sheened with a thick, clear fluid that could be tears or aqueous humor leaking from inside. I can see the iris, the pupil still perfectly dilated—what does that mean about the nervous system —and the white of the eye mottled with burst blood vessels.

Don't vomit. Don't you fucking dare vomit. Not now. Not with this goddamn eye staring at you.

I force myself to focus—on the objective. The evidence. What the hell happened here?

The scene in front of me tells a story of catastrophic violence. The question is—was this a murder or a murder-suicide? Or something else entirely? Something I'm not ready to acknowledge.

There's a third option. A terrifying possibility.

What if there wasn't a second person? What if the person who attacked you was you?

README history 1 version

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

  1. 2026-05-21Duplicate from DavidAU/Mistral-Nemo-Inst-2407-12B-Thinking-Uncensored-HERETIC...4bbad9713.4 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