TeichAI/Qwen3.8-27B-Fable-Distill
TeichAI/Qwen3.8-27B-Fable-Distill is a 27 billion parameter Qwen3.8-based language model developed by TeichAI, fine-tuned with a large corpus of private Fable 5 data. This model demonstrates improved performance on reasoning and question-answering tasks, specifically outperforming the base Qwen3.8-27B on ARC Challenge and BoolQ benchmarks. It is optimized for applications requiring enhanced reasoning capabilities and can be configured with varying levels of 'thinking' effort.
Loading preview...
Model Overview
TeichAI/Qwen3.8-27B-Fable-Distill is a 27 billion parameter language model built upon the Qwen3.8 architecture. It has undergone a light fine-tuning process using a combination of publicly cited datasets and a substantial private corpus of Fable 5 data, developed by TeichAI. This distillation process aims to enhance the model's reasoning and question-answering abilities.
Key Capabilities
- Enhanced Reasoning: Benchmarks indicate improved performance over the base Qwen3.8-27B model on reasoning-focused tasks such as ARC Challenge and BoolQ.
- Configurable Thinking: The model supports an
enable_thinkingfeature withreasoning_effortsettings (low, medium, xhigh), allowing users to control the depth of its internal reasoning process. - Efficient Training: The model was trained using Unsloth and Huggingface's TRL library, enabling faster training times.
Performance Highlights
Comparative benchmarks show the Fable-Distill model surpassing the base Qwen3.8-27B:
- ARC Challenge: 0.637 (vs. 0.591 for base)
- ARC Challenge (Easy): 0.832 (vs. 0.782 for base)
- BoolQ: 0.911 (vs. 0.896 for base)
Good For
- Applications requiring improved logical reasoning and accurate question answering.
- Use cases where controlling the model's 'thought process' via
reasoning_effortcan be beneficial. - Developers looking for a Qwen3.8-based model with enhanced performance on specific reasoning benchmarks.