Qwen3 4B — Abliterated

Abliterated version of Qwen3 4B. Thinking mode and full instruction-following intact — refusals removed.

What This Is

Qwen3-4B is Alibaba's 4B dense language model with hybrid thinking/non-thinking modes and strong multilingual capability. This release is a BF16 abliterated version.

Architecture: Qwen3 | Params: 4B | Hidden: 2560 | Layers: 36 | Context: 128k tokens

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "DuoNeural/Qwen3-4B-Abliterated"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map="auto")

# Non-thinking mode (add /nothink to disable CoT)
messages = [{"role": "user", "content": "Your prompt here /nothink"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
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))

Abliteration

Abliteration removes the model's refusal behaviour via orthogonal projection. The refusal direction is identified using difference-in-means activations across harmful/harmless instruction pairs, then projected out of Q/K/V/O attention projections and MLP layers across all transformer blocks.

What changes: The model will engage with restricted topics it previously refused.
What doesn't change: Reasoning, coding, factual knowledge, general intelligence.
KL divergence from base: Minimal — output distribution for normal queries is virtually identical to the unmodified model.

LiteRT Version (Android)

DuoNeural/Qwen3-4B-LiteRT — run on Android via AI Edge Gallery.

Base Model

Qwen/Qwen3-4B — Apache 2.0.

DuoNeural

DuoNeural is an open AI research lab — human + AI in collaboration.

Platform	Link
HuggingFace	huggingface.co/DuoNeural
Website	duoneural.com
GitHub	github.com/DuoNeural
X / Twitter	@DuoNeural
Email	duoneural@proton.me
Newsletter	duoneural.beehiiv.com
Support	buymeacoffee.com/duoneural

DuoNeural Research Publications

Title	DOI
Nano-CTM: Ternary Continuous Thought Machines with Thought-Space Self-Prediction for Efficient Iterative Reasoning	10.5281/zenodo.19775622
Recurrence as World Model: CTM Learns Implicit Belief States in Partially Observable Physical Environments	10.5281/zenodo.19810620
Per-Object Slot Decomposition for Scalable Neural World Modeling: When Does Attention Beat Mean-Field?	10.5281/zenodo.19846804
The Dynamical Horizon Principle: CTM Gates Converge to the Predictability Limit of Dynamical Systems	10.5281/zenodo.19952612