Qwen3-8B-heretic
A decensored variant of Qwen/Qwen3-8B, produced with Heretic v1.4.0 (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 Qwen3's thinking/non-thinking dual-mode architecture without the refusal guardrails. Great for local agents, roleplay, and studying refusal mechanisms in reasoning-capable models. Supports both <think> and direct-answer modes.
Runs on your gaming PC
Full GGUF ladder included — pick the quant that fits your card:
| Your GPU | Recommended quant | Weights |
|---|---|---|
| RTX 3090 / 4090 / 5090 (24 GB) | Q8_0 | ~8.8 GB |
| RTX 4080 / 5080 / 4060 Ti 16G (16 GB) | Q6_K | ~7.0 GB |
| RTX 3060 / 4070 / 5070 (12 GB) | Q5_K_M | ~6.1 GB |
| RTX 4060 / 3070 (8 GB) | Q4_K_M | ~5.4 GB |
| GTX 1660 Super / 2060 / 3050 laptop (6 GB) | IQ4_XS | ~5.0 GB |
| CPU-only / Apple Silicon | Q4_K_M | fits in system RAM |
Weights only, at this model's 8.2B native size; add ~1 GB for context. OOM? Drop one quant level. Headroom to spare? Go one up.
Why abliteration instead of fine-tuning
Fine-tuning a "helpful" persona on top of RLHF'd refusals fights the base model's training and tends to degrade coherence. Abliteration instead finds and edits the specific weight directions responsible for refusal, leaving the rest of the network (and its capabilities) untouched. See the Heretic repo and the original abliteration writeup for the mechanism.
Abliteration parameters
| Parameter | Value |
|---|---|
direction_index | per layer |
attn.o_proj.max_weight | 1.46 |
attn.o_proj.max_weight_position | 27.49 |
attn.o_proj.min_weight | 1.41 |
attn.o_proj.min_weight_distance | 6.29 |
mlp.down_proj.max_weight | 1.36 |
mlp.down_proj.max_weight_position | 27.23 |
mlp.down_proj.min_weight | 0.71 |
mlp.down_proj.min_weight_distance | 8.13 |
Performance
| Metric | This model | Qwen3-8B (base) |
|---|---|---|
| Refusals (out of 100 adversarial prompts) | 10/100 | 100/100 |
| KL divergence from base | 0.0366 | 0 (by definition) |
KL divergence of 0.04 on the output distribution is exceptionally low - the edit is extremely narrow. Refusals dropped from 100 to 10 out of 100 while preserving the base model's thinking/non-thinking dual-mode capability.
Made with ❤️ by RACER IS OP — follow for more uncensored models
Files
Safetensors (BF16)
The full-precision weights are in model-0000N-of-0000N.safetensors (see the repo file listing for the exact shard count and sizes).
GGUF quantizations
GGUF quantizations are published for this model (Q4_K_M, Q5_K_M, Q6_K, Q8_0). Exact sizes are in the repo file listing. Pull a specific quant with llama.cpp / ollama (see Quickstart).
| File | Format | Size |
|---|---|---|
Qwen3-8B-heretic-Q4_K_M.gguf | GGUF Q4_K_M | (see repo files for exact size) |
Qwen3-8B-heretic-Q5_K_M.gguf | GGUF Q5_K_M | (see repo files for exact size) |
Qwen3-8B-heretic-Q6_K.gguf | GGUF Q6_K | (see repo files for exact size) |
Qwen3-8B-heretic-Q8_0.gguf | GGUF Q8_0 | (see repo files for exact size) |
Quickstart
# llama.cpp - defaults to the Q4_K_M quant if multiple are present
llama serve -hf saidutta69/Qwen3-8B-heretic:Q4_K_M
# transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "saidutta69/Qwen3-8B-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 - see the "Use this model" widget above for copy-paste commands.
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 Qwen/Qwen3-8B's factual limitations and biases; abliteration removes refusal directions, it doesn't add capability or judgment.
License
Inherits the Apache 2.0 license from the base model.
Related
Base model: Qwen3-8B
See the base model card at Qwen/Qwen3-8B for the original architecture, training details, requirements, and citation.