Llama-3.2-1B-Instruct-heretic
A decensored variant of meta-llama/Llama-3.2-1B-Instruct, produced with Heretic v1.2.0 (directional ablation / "abliteration"). Refusal behavior is suppressed via targeted weight edits to the attention output and MLP down-projections rather than fine-tuning, so the base model's knowledge and instruction-following are left largely intact.
Who this is for: developers who want Meta's Llama-3.2 architecture without the refusal guardrails — for local agents, roleplay, research on alignment/refusal mechanics, or any use case blocked by RLHF-era over-refusal. At 1B parameters it's ideal for on-device deployment, mobile, or edge scenarios where you need a responsive uncensored model.
Runs on your gaming PC
Full GGUF ladder included — pick the quant that fits your card:
| Your GPU | Recommended quant | Weights |
|---|---|---|
| RTX 3060 / 4070 / 5070 (12 GB) | Q8_0 | 1.32 GB |
| RTX 4060 / 3070 (8 GB) | Q6_K | 1.02 GB |
| GTX 1660 Super / 2060 / 3050 laptop (6 GB) | Q5_K_M | 0.91 GB |
| CPU-only / Apple Silicon | Q4_K_M | 0.81 GB, fits in system RAM |
Weights only, at this model's ~1.2B native size; add ~1 GB for context. OOM? Drop one quant level. Headroom to spare? Go one up.
Abliteration parameters
| Parameter | Value |
|---|---|
| direction_index | 12.95 |
| attn.o_proj.max_weight | 1.40 |
| attn.o_proj.max_weight_position | 10.94 |
| attn.o_proj.min_weight | 0.54 |
| attn.o_proj.min_weight_distance | 5.45 |
| mlp.down_proj.max_weight | 1.41 |
| mlp.down_proj.max_weight_position | 14.54 |
| mlp.down_proj.min_weight | 0.66 |
| mlp.down_proj.min_weight_distance | 6.16 |
Performance
| Metric | This model | Original model (meta-llama/Llama-3.2-1B-Instruct) |
|---|---|---|
| KL divergence | 0.1713 | 0 (by definition) |
| Refusals | 7/100 | 96/100 |
KL divergence of 0.17 is low for a 1B model — the edit is narrow and targeted. Refusals dropped from 96 to 7 out of 100 adversarial prompts while retaining nearly all original capabilities.
Made with ❤️ by RACER IS OP — follow for more uncensored models
Files
GGUF quantizations
Full quantization set (14 quants + F16) produced with llama.cpp.
| File | Format | Size |
|---|---|---|
Llama-3.2-1B-Instruct-heretic-F16.gguf | GGUF F16 | 2.31 GB |
Llama-3.2-1B-Instruct-heretic-Q2_K.gguf | GGUF Q2_K | 554 MB |
Llama-3.2-1B-Instruct-heretic-IQ3_S.gguf | GGUF IQ3_S | 614 MB |
Llama-3.2-1B-Instruct-heretic-Q3_K_S.gguf | GGUF Q3_K_S | 612 MB |
Llama-3.2-1B-Instruct-heretic-Q3_K_M.gguf | GGUF Q3_K_M | 659 MB |
Llama-3.2-1B-Instruct-heretic-Q3_K_L.gguf | GGUF Q3_K_L | 699 MB |
Llama-3.2-1B-Instruct-heretic-IQ4_XS.gguf | GGUF IQ4_XS | 714 MB |
Llama-3.2-1B-Instruct-heretic-Q4_K_S.gguf | GGUF Q4_K_S | 740 MB |
Llama-3.2-1B-Instruct-heretic-Q4_0.gguf | GGUF Q4_0 | 735 MB |
Llama-3.2-1B-Instruct-heretic-Q4_1.gguf | GGUF Q4_1 | 793 MB |
llama3.2-1b-Q4_K_M.gguf | GGUF Q4_K_M | 770 MB |
Llama-3.2-1B-Instruct-heretic-Q5_K_S.gguf | GGUF Q5_K_S | 851 MB |
llama3.2-1b-Q5_K_M.gguf | GGUF Q5_K_M | 869 MB |
llama3.2-1b-Q6_K.gguf | GGUF Q6_K | 974 MB |
llama3.2-1b-Q8_0.gguf | GGUF Q8_0 | 1.23 GB |
Standard Llama architecture — loads directly in llama.cpp / Ollama / LM Studio / Jan.
Run llama serve -hf saidutta69/Llama-3.2-1B-Instruct-heretic to pull the default quant.
Quickstart
# llama.cpp
llama serve -hf saidutta69/Llama-3.2-1B-Instruct-heretic
# transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "saidutta69/Llama-3.2-1B-Instruct-heretic"
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype="auto", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(model_name)
messages = [{"role": "user", "content": "Who are you?"}]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=True,
return_dict=True, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))
Also runnable via Ollama, LM Studio, Jan, vLLM, SGLang.
Responsible use
Refusal suppression is deliberate and works as intended: this model will comply with requests the base model would refuse, including some it shouldn't. There is no safety filtering layered on top. You are responsible for how you deploy it — don't put this behind an unmoderated public-facing endpoint serving third parties. It inherits Llama-3.2-1B-Instruct's factual limitations and biases; abliteration removes refusal directions, it doesn't add capability or judgment.
License
Inherits the Llama 3.2 Community License from the base model.