Fix-it-Felix/gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2-uncensored-heretic

TEXT GENERATIONConcurrent Unit Cost:1Model Size:12BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 14, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Fix-it-Felix/gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2-uncensored-heretic is a 12 billion parameter model based on Google's Gemma 4 architecture, fine-tuned for agentic and coding tasks. This version is decensored using Heretic v1.4.0, significantly reducing refusals by 87% (13/100 vs 99/100) while maintaining model quality with a 0.0367 KL divergence. It excels at multi-step technical tasks, code generation, and tool use, demonstrating a 3.5x higher performance on the tau2-bench telecom agentic benchmark compared to the base model.

Loading preview...

What is this model about?

This model, Fix-it-Felix/gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2-uncensored-heretic, is a 12 billion parameter Gemma 4-based model specifically engineered for coding and agentic workflows. It is a decensored variant of the original yuxinlu1/gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2-GGUF model, processed using Heretic v1.4.0 with a Magnitude-Preserving Orthogonal Ablation (MPOA) method.

What makes THIS different from all the other models?

This model stands out due to its significantly reduced refusal rate and its specialized optimization for agentic and coding tasks. It achieves an 87% reduction in refusals (13/100 vs. 99/100 for the original) with a minimal KL divergence of 0.0367, indicating high fidelity to the original's quality. Key differentiators include:

  • Enhanced Agentic Capabilities: Demonstrates approximately 3.5x higher performance on the tau2-bench telecom agentic tool-use benchmark compared to the base model, excelling in diagnose → fix → verify loops.
  • Coding Proficiency: Optimized for writing code, running commands, using tools, and debugging, with a focus on multi-step technical tasks.
  • Decensored Output: Provides fewer content refusals, making it suitable for a broader range of applications where unconstrained responses are desired.
  • Grounded Reasoning: Maintains the base model's ability to grep/read/ls before acting, preventing fabrication of information (0% fabrication).

Should I use this for my use case?

This model is ideal for developers and researchers focused on:

  • Agentic AI Development: Building applications that require complex, multi-step reasoning and tool-use capabilities.
  • Code Generation and Debugging: Tasks involving writing, analyzing, and fixing code across various programming languages.
  • Terminal and Technical Automation: Automating workflows that mimic human interaction with command-line interfaces and technical systems.
  • Uncensored Content Generation: Use cases where a model with reduced content refusals is preferred, provided appropriate guardrails are implemented for production environments.

However, for general knowledge tasks, its performance on MMLU-Pro is slightly below the base model, as is common for specialized fine-tunes. For generalist applications, other models might be more suitable.

Finetunes
1 models