KevinBytes Logo
LLM Treeof Life

LLaMA 4 Scout Abliterated

Community (various contributors) · May 2025

● activeOpen Weightsparse moetext
Parameters81B (17B active × 16 experts)
Context Window10M tokens

Why It Matters

Demonstrated that abliteration techniques scale to frontier-class models, producing an uncensored model that rivals commercial offerings in raw capability.

Description

Widely considered the most capable uncensored model available. Uses a technique called abliteration — a mathematical method that identifies the specific 'direction' in the model's internal representations responsible for refusal responses, then surgically removes it without retraining. Built on Meta's LLaMA 4 Scout, which uses a Mixture-of-Experts architecture (only 17B of its 81B parameters activate for any given input, making it efficient despite its size).

Key Innovations

Abliteration
AbliterationRemoving safety guardrails from a model through targeted fine-tuning or weight manipulation. Controversial but popular in open-source community.
MoE
MoEArchitecture where only a fraction of the model's parameters are active for each input, allowing massive scale with lower compute.
Open Weight
Open WeightModel weights are publicly released but training data/code may not be. Enables fine-tuning but not full reproduction.

Family Tree

Built On

LLaMA 4

Lineage

LLaMA→LLaMA 2→LLaMA 3→LLaMA 3.1→LLaMA 3.2→LLaMA 3.3→LLaMA 4→LLaMA 4 Scout Abliterated
PreviousDeepSeek R1 Uncensored