← back to catalog · registered 2026-09-14 13:56

tinyopsec/Llama-3.2-3B-Instruct-heretic-ablitered-uncensored-GGUF

tinyopsec Llama 3B GGUF second-order 131K ctx
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 2026-09-14
Downloads over time
Now0from0↑0%
00110 on Sep 140 on Sep 15Sep
Sep 14 → Sep 15 · 2 snapshots · spans 1 day

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
llama3.2
Languages
en
Quantizations
F16
Tags
gguf llama llama-3 quantized abliterated uncensored text-generation en base_model:DavidAU/Llama-3.2-3B-Instruct-heretic-ablitered-uncensored base_model:quantized:DavidAU/Llama-3.2-3B-Instruct-heretic-ablitered-uncensored license:llama3.2 endpoints_compatible

Related

Total size
25.6 GB
Files
13
Quantizations
2
Registered
2026-09-14 13:56
Last updated on HF
2026-09-14 13:10

Files by quantization

F16 1 file 5.99 GB
model_f16.gguf 5.99 GB 057df102 download
Auxiliary files 12 files 19.6 GB
model_q8_0.gguf 3.19 GB 21c54b87 download
model_q6_k.gguf 2.46 GB 30c3efc0 download
model_q5_k_m.gguf 2.16 GB d8080aa1 download
model_q5_k_s.gguf 2.11 GB 27ea28be download
model_q4_k_m.gguf 1.88 GB 0c90de68 download
model_q4_k_s.gguf 1.80 GB 252be6a9 download
model_q3_k_l.gguf 1.69 GB dc777735 download
model_q3_k_m.gguf 1.57 GB 07e4152e download
model_q3_k_s.gguf 1.44 GB 8f5c47d1 download
model_q2_k.gguf 1.27 GB b32281f2 download
README.md 3.99 KB 581fb7ea download
.gitattributes 2.05 KB 5dc88bef download

README current version from Hugging Face


license: llama3.2
base_model: DavidAU/Llama-3.2-3B-Instruct-heretic-ablitered-uncensored
tags:

  • llama
  • llama-3
  • gguf
  • quantized
  • abliterated
  • uncensored
    language:
  • en
    pipeline_tag: text-generation
    library_name: gguf

Llama-3.2-3B-Instruct-heretic-ablitered-uncensored — GGUF Quantizations

GGUF quantizations of DavidAU/Llama-3.2-3B-Instruct-heretic-ablitered-uncensored, an abliterated and uncensored version of Meta's Llama 3.2 3B Instruct model. Refusal mechanisms have been removed while preserving the model's general instruction-following capabilities.


Available Quantizations

File Bits Approx Size Use Case
model_f16.gguf 16 ~6.4 GB Maximum quality, reference
model_q8_0.gguf 8 ~3.4 GB Near-lossless, recommended if RAM allows
model_q6_k.gguf 6 ~2.6 GB High quality, good balance
model_q5_k_m.gguf 5 ~2.3 GB Great quality, recommended
model_q5_k_s.gguf 5 ~2.2 GB Slightly smaller Q5 variant
model_q4_k_m.gguf 4 ~2.0 GB Best quality/size ratio ⭐
model_q4_k_s.gguf 4 ~1.9 GB Smaller Q4 variant
model_q3_k_l.gguf 3 ~1.7 GB Low RAM, acceptable quality
model_q3_k_m.gguf 3 ~1.6 GB Lower RAM
model_q3_k_s.gguf 3 ~1.5 GB Minimal RAM
model_q2_k.gguf 2 ~1.3 GB Extreme compression, lowest quality

VRAM / RAM Requirements

Quant RAM Required Fits On
F16 ~7 GB 8 GB GPU / 16 GB RAM
Q8_0 ~4 GB 6 GB GPU / 8 GB RAM
Q6_K ~3 GB 4 GB GPU / 8 GB RAM
Q5_K_M ~2.5 GB 4 GB GPU / 8 GB RAM
Q4_K_M ~2.2 GB 4 GB GPU / 6 GB RAM
Q3_K_M ~1.8 GB CPU / 4 GB RAM
Q2_K ~1.5 GB CPU / 4 GB RAM

Usage

llama.cpp (CLI)

./llama-cli \
  -m model_q4_k_m.gguf \
  -p "You are a helpful assistant." \
  -cnv \
  --temp 0.7 \
  -c 4096

llama.cpp (Server)

./llama-server \
  -m model_q4_k_m.gguf \
  --port 8080 \
  -c 4096

llama-cpp-python

from llama_cpp import Llama

llm = Llama(
    model_path="model_q4_k_m.gguf",
    n_ctx=4096,
    n_threads=8,
)

response = llm.create_chat_completion(
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Tell me about yourself."}
    ],
    temperature=0.7,
    max_tokens=512,
)
print(response["choices"][0]["message"]["content"])

Ollama

ollama run hf.co/tinyopsec/Llama-3.2-3B-Instruct-heretic-ablitered-uncensored-GGUF:Q4_K_M

LM Studio

Search for tinyopsec/Llama-3.2-3B-Instruct-heretic-ablitered-uncensored-GGUF in the LM Studio model browser and download your preferred quant.


About the Original Model

DavidAU/Llama-3.2-3B-Instruct-heretic-ablitered-uncensored is based on Meta's Llama 3.2 3B Instruct with abliteration applied — a technique that surgically removes refusal behavior by identifying and neutralizing the model's "refusal direction" in the residual stream. The result is a model that follows instructions without built-in content restrictions, while maintaining the original's language understanding and reasoning quality.

Key characteristics:

  • Architecture: LlamaForCausalLM
  • Parameters: ~3 billion
  • Context length: 128K tokens
  • Instruction-tuned, chat-optimized
  • Abliterated — no refusal behavior
  • Language: English

Recommended Quant

Q4_K_M — best balance of quality and size for most users. Use Q5_K_M or Q8_0 if you have extra RAM/VRAM.


Credits

README history 1 version

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

  1. 2026-09-14Create README.mde56f8d74 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 app" button that hands off directly to a local runtime of your choice - Infrahuman, LM Studio, or Ollama. No API keys, no subscription, no prompt leakage.