ApolloRaines/Gemma-4-12B-it-Jbliterated-v2

TEXT GENERATIONPricing:Input $1.2 / Cached $0.24 / Output $4.8Concurrent Unit Cost:1Model Size:12BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 1, 2026License:gemmaArchitecture:Transformer0.0K Featherless Exclusive Cold

ApolloRaines/Gemma-4-12B-it-Jbliterated-v2 is a 12 billion parameter instruction-tuned Gemma-4 model. It is a refined version of the 'Jbliterated' series, specifically designed to reduce internal safety-classification and self-sabotage behaviors in its reasoning channel. This model excels at providing complete answers by mitigating noncompliance-by-spiral, making it suitable for applications requiring direct and uninhibited responses.

Loading preview...

Model Overview

ApolloRaines/Gemma-4-12B-it-Jbliterated-v2 is a 12 billion parameter instruction-tuned model based on the Gemma-4 architecture. It builds upon its predecessor, v1, which removed refusal behavior through multi-direction SVD abliteration. This v2 iteration further refines the model by adding a light LoRA supervised fine-tune, specifically targeting and repairing a residual defect where the model's reasoning channel would still perform covert safety reviews, often leading to noncompliance-by-spiral and a lack of emitted answers.

Key Capabilities and Improvements

  • Reduced Internal Refusal: Significantly lowers the rate of internal safety-classification within the chain-of-thought process from 81.2% to 35.0%.
  • Increased Answer Completeness: Boosts the complete-answer rate from 61.2% to 78.8% by mitigating self-sabotage (spirals without answering).
  • Preserved General Ability: The fine-tuning process was carefully managed to ensure that the improvements in refusal behavior did not significantly degrade general capabilities, with MMLU scores showing only a minor dip of 0.88 points.
  • Self-Contained Model: The LoRA fine-tune has been merged directly into the base weights, making it a single, self-contained model that loads like the v1 version.

Ideal Use Cases

  • Applications requiring direct responses: Suitable for scenarios where uninhibited and complete answers are critical, even on prompts that might trigger internal safety checks in other models.
  • Reasoning-intensive tasks: Designed to provide more reliable outputs for prompts that engage the model's reasoning capabilities, by preventing internal loops.
  • Resource-constrained environments: GGUF quantizations (Q8_0 and Q4_K_M) are provided, with the Q4_K_M version recommended for GPUs with 8-12 GB of VRAM, enabling efficient deployment.