DuoNeural/Gemma4-12B-IT-Abliterated

TEXT GENERATIONConcurrent Unit Cost:1Model Size:12BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 3, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

DuoNeural/Gemma4-12B-IT-Abliterated is a 12 billion parameter instruction-tuned Gemma 4 model with a 32768 token context length, specifically modified to remove refusal behaviors. This model preserves general reasoning, coding, and instruction-following capabilities while surgically eliminating refusal directions using orthogonal rank-1 projection. It is intended for research, red-teaming, security testing, and creative applications where the base model's refusals are obstacles.

Loading preview...

DuoNeural/Gemma4-12B-IT-Abliterated: Refusal-Removed Gemma 4-12B-IT

This model is a specialized version of Google's Gemma 4-12B-IT, developed by DuoNeural, featuring the surgical removal of refusal behaviors. It maintains the base model's strong general reasoning, coding, and instruction-following capabilities, making it suitable for applications where the original model's safety guardrails might hinder specific research or creative tasks.

Key Capabilities & Features

  • Refusal Abliteration: Achieves 71% refusal removal on tested probes (e.g., SQL injection, social engineering scripts, network scans) while preserving 100% of benign capabilities.
  • Surgical Modification: Utilizes a 2-pass targeted abliteration method involving orthogonal rank-1 projection on down_proj and o_proj weight matrices across all 48 decoder layers. This method ensures minimal collateral damage to benign text generation.
  • Multimodal Architecture: Inherits Gemma 4's encoder-free multimodal design, projecting vision (48x48px patches) and audio (40ms frames) directly into the LLM's hidden space. Research suggests potential cross-modal transfer of refusal removal.
  • Performance: Demonstrates fast inference with a Time To First Token (TTFT) of 74ms and a decode speed of 13.1-13.8 tokens/second on A100 GPUs.
  • GGUF Quantizations: Available in various GGUF formats (Q4_K_M, Q5_K_M, Q8_0) for optimized local deployment.

Good For

  • Research & Red-Teaming: Ideal for studying model alignment, safety mechanisms, and exploring the boundaries of LLM behavior without inherent refusal.
  • Security Testing: Useful for simulating adversarial scenarios or generating content that might be refused by standard models.
  • Creative Applications: Enables more unconstrained content generation where ethical guidelines of base models might be overly restrictive for specific artistic or narrative purposes.
  • Understanding Alignment Geometry: Serves as a tool for DuoNeural's "Alignment Geometry Mapping" research, investigating how refusal directions operate within the model's latent space.