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

TEXT GENERATIONPricing:Input $1.2 / Cached $0.24 / Output $4.8Concurrent 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 model, fine-tuned by Richard Young, specifically optimized for verifiable Python coding tasks. It features a 256K token context length and is a decensored version of its base model, exhibiting significantly reduced refusals. This model excels at generating clean, runnable Python solutions by reasoning through edge cases and complexity.

Loading preview...

Model Overview

This model, richardyoung/gemma-4-12B-coder-fable5-composer2.5-v1-heretic, is a 12 billion parameter Gemma 4 variant, fine-tuned by Richard Young. It is a decensored version of yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1, created using Heretic v1.4.0, and demonstrates significantly reduced refusal rates (7/100 compared to 100/100 for the original).

Key Capabilities

  • Specialized Python Coding: Fine-tuned on verifiable Python coding data, it excels at generating clean, runnable solutions for algorithmic and function-level problems.
  • Chain-of-Thought Reasoning: Reasons openly about edge cases, complexity, and approach before writing code, leveraging a distillation of Composer 2.5 and Fable 5 CoT sources.
  • Extended Context Length: Features a corrected max_position_embeddings of 256K tokens, allowing for processing of longer code snippets and problem descriptions.
  • Reduced Refusals: Designed with task-focused training and no safety hedging, resulting in fewer refusals compared to its base model.
  • Full Precision Weights: Provided as un-quantized safetensors (bf16), suitable for custom quantizations (GGUF, MLX, AWQ, GPTQ) or further fine-tuning.

Ideal Use Cases

  • Code Generation: Best suited for generating Python code, particularly for algorithmic and deterministic testing scenarios.
  • Fine-tuning Base: Serves as a clean base for further LoRA or continued training for specific coding domains.
  • Custom Quantization: Developers can use the full-precision weights to create custom quantized versions for various deployment environments.

Limitations

  • Specialized Focus: Primarily optimized for Python and algorithmic coding; general knowledge facts may require double-checking.
  • English-Centric: Performance is optimized for English language inputs.
  • Not Safety-Aligned: Due to reduced safety hedging, users should implement their own guardrails for production environments.

Finetunes
1 models