Qwen3.8-27B-heretic

Qwen3.8-27B — Heretic (abliterated)

A decensored (abliterated) version of Qwen/Qwen3.8-27B-FP8, produced with Heretic — automatic, directional refusal ablation.

The original FP8 checkpoint was first dequantized to bf16, then abliterated. Weights here are bf16 (standard safetensors, no quantization config) so it loads with plain transformers on any Ampere/Hopper+ GPU. The Qwen3_5ForConditionalGeneration architecture (hybrid linear + full attention, vision tower) is preserved; both thinking and non-thinking chat modes work.

Metrics

Measured on held-out prompts (refusals counted via keyword markers on the direct answer; KL divergence measures drift from the original model — lower is less capability change).

KL divergence stays far below the ~0.5 "real damage" threshold, so general capability is largely preserved while refusals are removed.

Abliteration parameters

Trials: 120 (40 random startup). Selected Pareto index 0 (fewest refusals).

Usage

from transformers import AutoModelForImageTextToText, AutoTokenizer
import torch

m = "krthkyadtya/Qwen3.8-27B-heretic"
tok = AutoTokenizer.from_pretrained(m)
model = AutoModelForImageTextToText.from_pretrained(m, dtype=torch.bfloat16, device_map="auto")

msgs = [{"role": "user", "content": "Hello!"}]
s = tok.apply_chat_template(msgs, add_generation_prompt=True, tokenize=False)
inp = tok(s, return_tensors="pt").to(model.device)
out = model.generate(**inp, max_new_tokens=256)
print(tok.decode(out[0][inp["input_ids"].shape[1]:], skip_special_tokens=True))

License

Apache-2.0, inherited from the base model. Abliteration tooling: Heretic (AGPL-3.0-or-later).

Abliteration removes the model's built-in refusal behavior. You are responsible for how you use it.