richardyoung/gemma-4-12B-coder-fable5-composer2.5-v1-heretic

TEXT GENERATIONConcurrent Unit Cost:1Model Size:12BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 24, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The richardyoung/gemma-4-12B-coder-fable5-composer2.5-v1-heretic is a 12 billion parameter Gemma 4-based language model, fine-tuned for verifiable Python coding tasks with a 32768 token context length. This model is a decensored version of yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1, created using Heretic v1.4.0. It excels at generating clean, runnable Python solutions by reasoning through edge cases and complexity. Its primary strength lies in algorithmic and function-level Python coding, making it suitable for developers requiring robust code generation.

Loading preview...

Model Overview

This model, richardyoung/gemma-4-12B-coder-fable5-composer2.5-v1-heretic, is a 12 billion parameter Gemma 4-based language model, specifically fine-tuned for verifiable Python coding. It is a decensored variant of the original yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1, processed with Heretic v1.4.0. The model features a corrected context length of 256K tokens, addressing an upstream Gemma 4 metadata bug.

Key Capabilities

  • Verifiable Python Coding: Optimized for generating Python code that passes tests, reasoning through edge cases and complexity.
  • Reduced Refusals: Designed with task-focused training and no safety hedging, resulting in fewer refusals compared to the base model.
  • Full Precision Weights: Provided as un-quantized safetensors (bf16) for custom quantization, further fine-tuning, or direct use with transformers.
  • Agentic Behavior: Version 2, currently in benchmarking, significantly expands agentic and coding data for enhanced performance.

Training Data

The model was trained on a distillation of two chain-of-thought sources for verifiable Python coding tasks:

  • Composer 2.5: Genuine model-authored reasoning traces, with solutions verified against task tests.
  • Fable 5: Used to re-derive fresh, self-consistent CoT and correct solutions for problems Composer 2.5 initially failed, also gated on passing tests.

Good For

  • Developers needing a model for Python code generation, especially for algorithmic and function-level problems.
  • Users looking to create custom quantizations (GGUF, MLX, AWQ, GPTQ) or perform further fine-tuning (LoRA).
  • Applications requiring a model with reduced safety-alignment for specific coding tasks, with the understanding that guardrails may need to be added for production use.

Finetunes
1 models