enkato64bit/SuperGemma-4-12b-abliterated

TEXT GENERATIONPricing:Input $1.2 / Cached $0.24 / Output $4.8Concurrent Unit Cost:1Model Size:12BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 17, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

SuperGemma-4-12b-abliterated is a 12 billion parameter model derived from Google's Gemma-4-12B-it, developed by enkato64bit. This fused checkpoint integrates an "Abliteration pass" to reduce refusal behavior and "Supertune post-training" for enhanced instruction following, coding, and JSON/tool formatting. It significantly improves performance on coding benchmarks like HumanEval+ and MBPP+, making it suitable for direct task completion and technical applications.

Loading preview...

SuperGemma-4-12b-abliterated: Enhanced Gemma-4-12B-it

SuperGemma-4-12b-abliterated is a 12 billion parameter model based on Google's Gemma-4-12B-it, developed by enkato64bit. This model is a single, fused checkpoint that combines two post-training stages to deliver improved performance without requiring runtime adapters.

Key Enhancements and Capabilities

  • Abliteration Pass: This stage focuses on suppressing unnecessary refusal behavior, leading to more direct and compliant task completion.
  • Supertune Post-training: Targeted post-training enhances several critical areas:
    • Improved instruction following.
    • Stronger coding capabilities.
    • Better Korean technical answers.
    • Enhanced JSON and tool formatting.
    • Increased regression resistance.

Performance Benchmarks

Compared to the original Gemma4 12B instruction checkpoint, SuperGemma-4-12b-abliterated shows significant gains:

  • Overall public top-5 500: +20.8 delta (from 23.8 to 44.6)
  • HumanEval+: +28.0 delta (from 18.0 to 46.0)
  • MBPP+: +68.0 delta (from 13.0 to 81.0)
  • GPQA Diamond: +9.0 delta (from 10.0 to 19.0)
  • MMLU-Pro: +1.0 delta (from 17.0 to 18.0)

Internal validation also indicates a 0.0 blank response ratio and 0.0 hidden-thought leak ratio, with strong overall scores on Quickbench full20 (95.4) and Mega 103 (88.7).

Use Cases

This model is particularly well-suited for applications requiring:

  • Direct task completion: Optimized to reduce refusals and provide straightforward answers.
  • Code generation and understanding: Significant improvements on coding benchmarks make it effective for programming-related tasks.
  • Structured output: Enhanced JSON and tool formatting capabilities are beneficial for agentic workflows or data processing.
  • Technical Q&A: Especially noted for Korean technical answers.

Finetunes
1 models