AEON-7/Gemma-4-12B-it-AEON-Abliterated-K4-BF16
AEON-7/Gemma-4-12B-it-AEON-Abliterated-K4-BF16 is a 12 billion parameter bfloat16 model based on Google's Gemma-4-12B-it, developed by AEON-7. This model has undergone a K=4 multi-direction biprojection abliteration process to remove refusal behaviors while preserving core capabilities. It is optimized for fine-tuning and non-Blackwell hardware, offering capability-neutral performance compared to its base model.
Loading preview...
Model Overview
AEON-7/Gemma-4-12B-it-AEON-Abliterated-K4-BF16 is a 12 billion parameter bfloat16 model derived from google/gemma-4-12B-it through a K=4 multi-direction biprojection process. This "abliteration" technique effectively removes refusal behaviors, allowing the model to generate full responses to prompts it would typically decline, while maintaining its original capabilities.
Key Capabilities & Features
- Refusal-Removed: Generates full responses to prompts that the base model would refuse, often with a brief disclaimer.
- Capability-Preserving: Benchmarks show performance within ~1 percentage point of the base
google/gemma-4-12B-itacross MMLU, HumanEval, and IFEval. - Optimized for Specific Hardware: This BF16 variant is recommended for fine-tuning tasks or deployment on non-Blackwell hardware.
- High Concurrency Throughput: Achieves near-linear scaling in aggregate throughput, peaking at ~450-460 tokens/second at c=64 on DGX Spark GB10 with optimized vLLM.
- Technical Architecture: Features 48 decoder layers, 3840 hidden size, GQA attention, and a 262,144 token vocabulary.
Good for
- Fine-tuning: Provides a bit-exact weight anchor for further customization.
- Non-Blackwell Hardware: Ideal for systems not equipped with Blackwell architecture, where its FP8/NVFP4 siblings offer speed advantages.
- Applications Requiring Unfiltered Responses: Suitable for use cases where the base model's refusal behaviors are undesirable, provided appropriate downstream safety layers are implemented.