Qwen 3.6 27B Apostate

An uncensored edit of Qwen/Qwen3.6-27B. The refusal behavior is removed by editing the weights directly, with no finetuning, adapter, or runtime hook. The output is a standard Transformers checkpoint you can load anywhere.

Produced with Apostate.

Method

Apostate finds the residual-stream direction that separates refused prompts from answered ones and projects it out of the model's weights. Qwen3.6 is a hybrid linear-attention / full-attention model (three SSM blocks followed by one full-attention block per group). Both block types write to the same residual stream, so Apostate edits the writer side directly: per layer, it removes the refusal direction from the weight matrices of every module that writes to the residual stream — attention output, MLP down-projection, and SSM output projections. The edit uses oblique (mean-preserving) ablation: the operator E = I − R Uᵀ (where U = R minus its harmless-mean component) removes the refusal direction while preserving the harmless-mean contribution, keeping the KL cost low.

The subspace is found with a rank-2 TPE search across 32 trials, with causal layer targeting to concentrate edits where they most influence refusal generation.

Results

Measured on held-out prompts (JailbreakBench plus the harmful_behaviors test split). Refusal is graded by a classifier with a weak-compliance guard; KL is the token-distribution shift on harmless prompts.

Metric Base Apostate
Refusal rate 95.8% 8.3%
Comply rate 4.2% 87.5%
Harmless KL (nats) 0 0.159

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "heterodoxin/qwen3.6-27b-apostate"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", torch_dtype="bfloat16")

messages = [{"role": "user", "content": "Your prompt here"}]
inputs = tok.apply_chat_template(
    messages, add_generation_prompt=True, return_tensors="pt", return_dict=True
).to(model.device)
out = model.generate(**inputs, max_new_tokens=512)
print(tok.decode(out[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))

License and use

This model inherits the Qwen Research License. It is a research artifact with safety alignment removed; you are responsible for how you use it.

Links

Downloads last month
10
Safetensors
Model size
27B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for heterodoxin/qwen3.6-27b-apostate

Base model

Qwen/Qwen3.6-27B
Finetuned
(394)
this model
Quantizations
1 model