yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1

TEXT GENERATIONConcurrent Unit Cost:1Model Size:12BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 17, 2026License:apache-2.0Architecture:Transformer0.1K Open Weights Featherless Exclusive Cold

The yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1 is a 12 billion parameter Gemma 4 based language model, fine-tuned by yuxinlu1, specifically optimized for verifiable Python code generation and algorithmic problem-solving. It features an extended context length of 256K tokens and excels at reasoning through coding challenges to produce clean, runnable solutions. This model is ideal for developers requiring a robust coding assistant for Python-centric tasks.

Loading preview...

Model Overview

yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1 is a 12 billion parameter Gemma 4 model, meticulously fine-tuned by yuxinlu1 for Python code generation. This version is the full-precision safetensors master, providing the original weights for advanced users to create custom quantizations (GGUF, MLX, AWQ, GPTQ) or for further fine-tuning.

Key Capabilities

  • Verifiable Python Coding: Specialized in generating runnable Python code for algorithmic and function-level problems, with solutions verified against tests during training.
  • Extended Context Window: Features a corrected max_position_embeddings of 256K tokens, enabling processing of longer codebases and complex problem descriptions.
  • Reasoning in the Open: Designed to articulate its thought process (edge cases, complexity, approach) before providing a solution, enhancing transparency and debuggability.
  • Reduced Refusals: Task-focused training minimizes refusals compared to the base model, making it more direct in its responses.
  • Apache 2.0 License: Inherits the Apache 2.0 license from its Gemma 4 base, allowing free use, modification, and redistribution.

Training Methodology

The model was trained using a distillation approach from two verifiable Python coding chain-of-thought (CoT) sources:

  • Composer 2.5 Real CoT: Genuine model-authored reasoning traces, with solutions validated by execution against task tests.
  • Fable 5 Redo: For problems where Composer 2.5 failed, Fable 5 re-derived self-consistent CoT and correct solutions, also gated on passing tests. This synthetic data patched failures from the primary teacher.

Good For

  • Developers needing a powerful assistant for Python code generation and problem-solving.
  • Researchers looking for a clean base for further fine-tuning on coding-related tasks.
  • Users who want to create custom quantized versions of a highly capable coding model.

Finetunes
2 models