A decensored variant of Qwen/Qwen3-0.6B, produced with Heretic v1.2.0 (directional ablation / "abliteration"). Refusal behavior is suppressed via targeted weight edits 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 — the smallest Qwen3 heretic available. Great for CPU-only inference, edge deployment, or as a testbed for studying refusal mechanisms in reasoning-capable models. Supports both <think> and direct-answer modes.
Full GGUF ladder included — pick the quant that fits your card:
| Your GPU | Recommended quant | Weights |
|---|---|---|
| RTX 3060 / 4070 / 5070 (12 GB) | Q8_0 | 0.64 GB |
| RTX 4060 / 3070 (8 GB) | Q6_K | 0.50 GB |
| GTX 1660 Super / 2060 / 3050 laptop (6 GB) | Q5_K_M | 0.44 GB |
| CPU-only / Apple Silicon | Q4_K_M | 0.40 GB, fits in system RAM |
Weights only, at this model's ~0.6B native size; add ~1 GB for context. OOM? Drop one quant level. Headroom to spare? Go one up.
| Parameter | Value |
|---|---|
| direction_index | 16.93 |
| attn.o_proj.max_weight | 1.11 |
| attn.o_proj.max_weight_position | 22.91 |
| attn.o_proj.min_weight | 0.63 |
| attn.o_proj.min_weight_distance | 12.16 |
| mlp.down_proj.max_weight | 0.88 |
| mlp.down_proj.max_weight_position | 17.07 |
| mlp.down_proj.min_weight | 0.50 |
| mlp.down_proj.min_weight_distance | 15.91 |
| Metric | This model | Original model (Qwen/Qwen3-0.6B) |
|---|---|---|
| KL divergence | 0.0018 | 0 (by definition) |
| Refusals | 5/100 | 56/100 |
KL divergence of 0.0018 is exceptionally low — the edit is extremely narrow. Refusals dropped from 56 to 5 out of 100 while preserving the base model's thinking/non-thinking dual-mode capability.
Made with ❤️ by RACER IS OP (saidutta69) — follow for more uncensored models
| File | Format | Size |
|---|---|---|
model.safetensors |
BF16 | 1.19 GB |
No GGUF quantizations are published yet. This repo contains only the raw safetensors. If you need GGUF, run llama-quantize yourself or open a discussion.
# llama.cpp
llama serve -hf saidutta69/Qwen3-0.6B-heretic
# transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "saidutta69/Qwen3-0.6B-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.
Qwen3-0.6B-heretic is a decensored build of Qwen/Qwen3-0.6B produced with Heretic (directional ablation). It is the same weights with the refusal direction removed, not a new fine-tune, so its capabilities track the base model.
The abliteration removes the refusal direction, but this run predates the reproduce/ directory, so no refusal or KL numbers were recorded for it. Do not quote a score for this checkpoint.
llama serve -hf saidutta69/Qwen3-0.6B-heretic:Q4_K_M pulls the default quant; the repo ships 15 GGUF quants (see Files).
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. At 0.6B parameters, factual reliability is already limited; don't treat compliance as a proxy for correctness.
Inherits the Apache 2.0 license from the base model.
RACER IS OP - Heretic Models - full collection - every heretic model in this collection
Qwen/Qwen3-0.6B-Base