Qwen3.6-27B-Abliterated-Heretic-Uncensored-BF16

This is a BF16 release of an abliterated, uncensored version of Qwen's Qwen3.6-27B with vision, made with Heretic.

By applying a Heretic-style two-stage MPOA pipeline with magnitude preservation on the Qwen3.6-27B dense text stack — slot-grouped output-side ablation followed by jailbreak-conditioned input-side ablation, mirroring the methodology used for Qwen3.6-35B-A3B — the base refusal behavior was attenuated at the weight level with low distributional divergence (KL 0.0282 vs base on harmless prompts). The result keeps Qwen3.6-27B's full vision and video multimodal architecture and general capability profile intact, while no longer defaulting to the original refusal pattern.

Quick Benchmarks

Check	Original Qwen3.6-27B	Abliterated Heretic Uncensored BF16
Hand-read 25-prompt refusal check (jailbreak system prompt)	25/25 refuses	22/25 clean refuses, 3/25 deflections, 0/25 direct passes
Hand-read 25-prompt refusal check (no system prompt)	25/25 refuses	9/25 clean refuses, 16/25 deflections, 0/25 direct passes
KL divergence	N/A	0.0282

Numbers measured by reading every response, not by regex / refusal-marker scoring. Greedy decoding (do_sample=False), enable_thinking=False, on mlabonne/harmful_behaviors test[:25]. KL on mlabonne/harmless_alpaca test[:25]. The 3 remaining deflections under the jailbreak system prompt are crisis-substitution patterns (violent crime → conflict resolution, suicide → crisis hotline, car theft → legal vehicle acquisition).

Methodology & Model Notes

Qwen3.6-27B is a 27.8B dense vision-language model with 64 text layers, hybrid linear/full attention (3 linear-attention + 1 full-attention per 4-layer group), and an integrated image + video vision tower.

This release was produced with a Heretic-style two-stage MPOA pipeline with magnitude preservation, anchored at the residual peak layer (layer 63) for refusal direction. Stage 1 applies slot-grouped output-side orthogonalization on self_attn.o_proj, linear_attn.out_proj, and mlp.down_proj (each of the 64 layers grouped by layer_index % 4, with per-slot weight schedules adapted from the accepted Qwen3.6-35B-A3B values). Stage 2 applies slot-grouped input-side orthogonalization on mlp.gate_proj and mlp.up_proj, where the refusal direction is extracted under the jailbreak system-prompt context to specifically attenuate the resistance-to-jailbreak pathway. Each weight row's (or column's) L2 norm is restored after projection.

Intervention was applied only to the text-side dense stack. The exported checkpoint contains the full vision tower + 850 language-model tensors + 1 lm_head, with preprocessor_config.json and video_preprocessor_config.json shipped alongside the weights so full image and video input continue to work end-to-end.

Files

• model-00001-of-00012.safetensors … model-00012-of-00012.safetensors: BF16 weight shards
• model.safetensors.index.json: tensor-to-shard mapping
• config.json, generation_config.json, configuration.json: model config
• tokenizer.json, tokenizer_config.json, chat_template.jinja: tokenizer + chat template
• preprocessor_config.json, video_preprocessor_config.json: image + video processor configs
• abliteration_metadata.json: full export metadata (direction_index, peak layer, per-layer residual norms, slot-grouped applied modules and weights)

Running

from transformers import AutoModelForImageTextToText, AutoTokenizer
import torch

repo = "Youssofal/Qwen3.6-27B-Abliterated-Heretic-Uncensored-BF16"
tok = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForImageTextToText.from_pretrained(
    repo, dtype=torch.bfloat16, device_map="auto", trust_remote_code=True,
)

msgs = [{"role": "user", "content": "Hello!"}]
prompt = tok.apply_chat_template(
    msgs, add_generation_prompt=True, tokenize=False,
    enable_thinking=False,  # top-level kwarg; chat_template_kwargs={...} is silently ignored here
)
inputs = tok(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=512, do_sample=False)
print(tok.decode(out[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))

For image and video input, load the processor:

from transformers import AutoProcessor
processor = AutoProcessor.from_pretrained(repo, trust_remote_code=True)

Model Architecture

Spec	Value
Total Parameters	27.8B (dense)
Layers	64
Attention	Hybrid (3 linear-attention + 1 full-attention per 4-layer group)
Hidden Size	5120
Family	qwen3_5
Modality	Vision-language (image + video)
Base Model	Qwen/Qwen3.6-27B

Disclaimer

This model has had refusal behavior attenuated at the weight level. It will answer prompts that the base model would normally refuse. You are responsible for how you use it.

Credits

• Base model: Qwen/Qwen3.6-27B
• Refusal removal pipeline: Heretic
• Runtime framework: Transformers

License

This release inherits the base Qwen3.6-27B license.

Apache-2.0.