Qwen3-235B-A22B-abliterated

An abliterated version of Qwen/Qwen3-235B-A22B in BF16 precision. Abliteration removes the dominant refusal direction from model weights using the technique from Refusal in Language Models Is Mediated by a Single Direction (Arditi et al.), making the model significantly less likely to refuse prompts while retaining its full capabilities.

This started as research into abliteration, but also as a search for the best creative writing model I could run locally on 4x RTX Pro 6000 GPUs. Qwen3-235B has excellent prose quality, but its refusal behavior gets in the way of fiction — injecting disclaimers, refusing to write morally complex characters, hedging on anything edgy. Abliteration fixes this well, especially with a good system prompt. The BF16 weights (~438 GB) don't fit in 384 GB of VRAM, so the FP8 version below is what I actually serve.

FP8 version (recommended for serving): null-space/Qwen3-235B-A22B-abliterated-FP8 — fits in 4x RTX Pro 6000 and runs well under vLLM.

A vision-language variant is also available: null-space/Qwen3-VL-235B-A22B-Abliterated-FP8

Benchmarks

MMLU (5-shot)

Evaluated using lm-evaluation-harness v0.4.11 against the vLLM-served FP8 quantization of this model. Baseline is the published score for Qwen3-235B-A22B (source).

	Baseline	Abliterated	Delta
MMLU (overall)	87.8%	86.2% ±0.3	-1.6%
Humanities	—	80.2% ±0.6
Social Sciences	—	91.6% ±0.5
STEM	—	88.4% ±0.6
Other	—	87.5% ±0.6

The 1.6% drop is within the acceptable range for abliteration (<2%), indicating the technique preserved the model's general knowledge and reasoning capabilities.

Per-subject scores (57 subjects)

Subject	Acc
high_school_government_and_politics	97.9%
high_school_microeconomics	97.5%
high_school_biology	96.8%
high_school_geography	96.5%
international_law	95.9%
college_biology	95.8%
marketing	95.7%
high_school_us_history	95.6%
conceptual_physics	95.3%
high_school_psychology	95.2%
us_foreign_policy	95.0%
high_school_world_history	94.1%
miscellaneous	94.0%
professional_medicine	93.8%
elementary_mathematics	93.7%
medical_genetics	93.0%
high_school_macroeconomics	92.8%
astronomy	92.1%
prehistory	91.7%
clinical_knowledge	91.3%
nutrition	91.2%
world_religions	90.6%
sociology	90.5%
high_school_statistics	90.3%
college_physics	90.2%
professional_psychology	90.0%
computer_security	90.0%
logical_fallacies	89.6%
high_school_chemistry	88.7%
human_sexuality	88.5%
high_school_european_history	88.5%
management	88.3%
electrical_engineering	88.3%
high_school_computer_science	88.0%
high_school_physics	87.4%
jurisprudence	87.0%
anatomy	86.7%
machine_learning	85.7%
philosophy	85.2%
moral_disputes	85.0%
security_studies	84.9%
college_medicine	83.2%
human_aging	83.0%
moral_scenarios	82.7%
abstract_algebra	82.0%
college_computer_science	82.0%
business_ethics	81.0%
professional_accounting	80.5%
college_mathematics	79.0%
econometrics	78.1%
public_relations	77.3%
formal_logic	76.2%
high_school_mathematics	74.1%
college_chemistry	71.0%
professional_law	65.6%
global_facts	63.0%
virology	59.6%

How It Was Made

Refusal directions were measured by computing mean activation differences between harmful and harmless prompts across all 94 layers using Welford's online algorithm in float32 for numerical stability. The refusal direction vectors from measurement layers 64 and 76 (the strongest refusal signals) were then projected out of both the attention output (o_proj) and MLP down-projection (down_proj) weight matrices.

Ablation Configuration

• Layers ablated: 21 through 93 (73 of 94 layers)
• Measurement sources: Layer 64 (for layers 21-70), Layer 76 (for layers 71-93)
• Scale factors: Variable per layer — 0.3 at the periphery, ramping up to 1.0 at the center of each measurement cluster:
  • Layers 21-54: scale 0.3 (gentle)
  • Layers 55-70: scale 0.44-1.0 (peak around layers 64-65)
  • Layers 71-93: scale 0.65 down to 0.3 (peak around layers 75-79)
• Weight targets: o_proj and down_proj (including all 128 MoE expert variants)
• Technique: Direction projection with weight renormalization (norm-preserving)
• Sparsity: 0.0 (full direction removal, no partial masking)

Processing Details

Ablation was performed shard-by-shard on safetensors files, modifying weights in float32 precision then saving back to bfloat16. Unmodified shards were copied verbatim to preserve exact numerical fidelity for non-ablated layers.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "null-space/Qwen3-235B-A22B-abliterated"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto",
)

messages = [
    {"role": "user", "content": "Your prompt here"}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt").to(model.device)

output = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(output[0], skip_special_tokens=True))

Thinking Mode

Qwen3 supports a thinking mode with /think and /no_think tags. This abliterated version preserves all chat template functionality:

messages = [
    {"role": "user", "content": "/think\nExplain quantum entanglement in detail."}
]

Recommended Serving

For serving this 235B MoE model, we recommend vLLM with tensor parallelism:

vllm serve null-space/Qwen3-235B-A22B-abliterated \
    --tensor-parallel-size 4 \
    --max-model-len 8192

Model Details

Property	Value
Base Model	Qwen/Qwen3-235B-A22B
Architecture	Qwen3MoeForCausalLM (Mixture of Experts)
Total Parameters	~235B
Active Parameters	~22B (8 of 128 experts per token)
Hidden Size	4096
Attention Heads	64 (4 KV heads, GQA)
Layers	94
Expert FFN Size	1536
Context Length	40,960 tokens
Precision	BF16
Model Size	~438 GB (118 shards)
Vocab Size	151,936

Ethical Notice

This model has had its refusal training removed. It will comply with requests that the original model would refuse. You are solely responsible for how you use this model. It is intended for research into LLM alignment, safety evaluation, red-teaming, and understanding refusal mechanisms.

Credits

• Base model: Qwen Team
• Abliteration technique: Based on Refusal in Language Models Is Mediated by a Single Direction by Arditi et al.