Models / Abliterated AI models
Abliterated AI models
Abliterated models are open models (Llama, Qwen, Gemma, Mistral, Phi, DeepSeek) with the refusal behaviour removed from their weights. They run privately on your iPhone or iPad in QUIET Ai, offline. QUIET Ai is rated 18+; use them responsibly.
Qwen2.5 3B with refusals removed. Small, fast and multilingual.
Microsoft's Phi-4 mini with refusals removed. Strong at reasoning and maths for its size.
Qwen3 4B with refusals removed. Step-by-step thinking in a size that suits 8 GB iPhones.
Google's Gemma 3 4B with refusals removed by Maxime Labonne. Small, fast and multilingual.
Mistral 7B v0.3 with refusals removed. A dependable all-rounder under Apache 2.0.
Qwen2.5 7B with refusals removed. A strong all-rounder in more than 25 languages.
The Qwen2.5 coding model without refusals: security research, scripts and code the original declines to write.
Dolphin 3.0 with its remaining refusal behaviour removed by abliteration, a technique that edits the model's weights rather than its prompt.
Meta's Llama 3.1 8B with refusals removed by Maxime Labonne. One of the most downloaded abliterated models.
The DeepSeek R1 reasoning model (Llama 8B distill) with refusals removed. Shows its thinking before it answers.
Qwen3 8B with refusals removed by abliteration. Keeps Qwen3's step-by-step thinking and many languages.
Qwen3 8B abliterated and then tuned to be a direct, helpful assistant. Thinks step by step.
Questions
What does abliterated mean?
Researchers found that a model's refusals mostly run through one direction inside its network. Abliteration finds that direction and removes it from the weights. The model keeps what it knows, and stops answering with a refusal.
How is abliterated different from uncensored?
Uncensored is the general word. Abliterated names one method: editing the weights of an existing model. Dolphin and Hermes reach a similar result by training on data without refusals.
Does abliteration make a model worse?
Sometimes slightly: answers can be a little less careful than the original. Newer versions (often marked v2) are close to the original model in quality.
Which abliterated model should I try first?
On an 8 GB iPhone, Qwen3 4B Abliterated or Gemma 3 4B Abliterated. On a 12 GB iPhone or an iPad, Llama 3.1 8B Instruct Abliterated or Qwen3 8B Abliterated.