Overview
DavidAU/Gemma-3-27b-it-Gemini-Extreme-Deep-Reasoning is a 27 billion parameter Gemma-3 model, instruction-tuned by DavidAU using Unsloth. Its primary differentiator is the integration of "Gemini" type deep reasoning, achieved through fine-tuning with the "gemini-3-pro-preview-high-reasoning" dataset. This tuning enables the model to automatically generate detailed thought blocks for complex queries, significantly improving output quality even in standard instruct mode. It supports a 128k context length and maintains reasoning stability across a wide temperature range.
Key Capabilities
- Enhanced Deep Reasoning: Automatically activates multi-paragraph thought processes for complex problems, without requiring specific system prompts in most cases.
- Improved Output Quality: The fine-tuning generally enhances the quality of model responses, even when direct reasoning isn't explicitly triggered.
- Image Intelligence: Reasoning capabilities extend to image processing, improving the model's understanding and generation related to visual inputs.
- Flexible Activation: While automatic, users can explicitly prompt for thinking with "Think deeply:" or utilize specialized Jinja templates for always-on reasoning.
Benchmarks & Performance
Compared to a base Gemma model, this version shows notable improvements across several benchmarks:
- ARC Challenge: 0.577 (vs 0.557)
- Hellaswag: 0.784 (vs 0.533)
- Winogrande: 0.763 (vs 0.695)
Good For
- Applications requiring robust logical deduction and problem-solving.
- Tasks where detailed, multi-step reasoning is beneficial.
- Use cases involving image understanding and generation where enhanced "intelligence" is desired.
- Users seeking a Gemma-3 variant with significantly improved cognitive abilities beyond basic instruction following.