ThalisAI/Qwen3-32B-heretic

WARNING: this model was converted before I implemented a fix for thinking models, which led to a falsely low KL divergence value of 0.0000. That caused the Optuna optimizer to select poor trials that destroyed model coherence. I will be re-abliteratin this one.

Qwen3-32B-heretic

Abliterated (uncensored) version of Qwen/Qwen3-32B, created using Heretic and converted to GGUF.

Abliteration Quality

Lower refusals = fewer refused prompts. Lower KL divergence = closer to original model behavior.

Note: KL divergence of 0.0000 is expected for this model. Qwen3-32B uses a <think></think> response prefix, which means the first-token probability distribution on harmless prompts is identical before and after abliteration. The abliteration successfully removes refusal behavior (0/100 refusals) while leaving the model's harmless response behavior completely unchanged.

Available Quantizations

Usage with Ollama

ollama run hf.co/ThalisAI/Qwen3-32B-heretic:Q8_0
ollama run hf.co/ThalisAI/Qwen3-32B-heretic:Q6_K
ollama run hf.co/ThalisAI/Qwen3-32B-heretic:Q4_K_M

bf16 Weights

The full bf16 abliterated weights are available in the bf16/ subdirectory of this repository.

Usage with Transformers

The bf16 weights in the bf16/ subdirectory can be loaded directly with Transformers:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "ThalisAI/Qwen3-32B-heretic"
tokenizer = AutoTokenizer.from_pretrained(model_id, subfolder="bf16")
model = AutoModelForCausalLM.from_pretrained(
    model_id, subfolder="bf16", torch_dtype="auto", device_map="auto"
)

messages = [{"role": "user", "content": "Hello!"}]
text = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

About

This model was processed by the Apostate automated abliteration pipeline:

1. 1.The source model was loaded in bf16
2. 2.Heretic's optimization-based abliteration was applied to remove refusal behavior
3. 3.The merged model was converted to GGUF format using llama.cpp
4. 4.Multiple quantization levels were generated

The abliteration process uses directional ablation to remove the model's refusal directions while minimizing KL divergence from the original model's behavior on harmless prompts.