DavidAU/Gemma3-27B-it-vl-GLM-4.7-Uncensored-Heretic-Deep-Reasoning is a 27 billion parameter Gemma 3 instruction-tuned model, fine-tuned with the GLM 4.7 reasoning dataset via Unsloth. This model features 128k context and is fully uncensored, excelling in deep reasoning and image intelligence. It is optimized for direct, detailed responses across various tasks, including those requiring explicit content, and shows improved benchmarks over its non-thinking counterpart.
Loading preview...
Model Overview
DavidAU/Gemma3-27B-it-vl-GLM-4.7-Uncensored-Heretic-Deep-Reasoning is a 27 billion parameter Gemma 3 instruction-tuned model, enhanced with deep reasoning capabilities. It leverages the GLM 4.7 reasoning dataset and Unsloth for fine-tuning, resulting in a fully uncensored model designed to provide direct and detailed responses without refusal.
Key Capabilities & Features
- Deep Reasoning: Significantly improves image "intelligence" and general model operation, leading to enhanced output generation and benchmark performance.
- Uncensored Output: Designed to generate content exactly as requested, including explicit or sensitive topics, with minimal refusal. It may require specific directives (e.g., "use these words to swear") for highly graphic content.
- Extended Context: Features a 128k context window, allowing for processing longer inputs and maintaining coherence.
- Temperature Stability: Reasoning capabilities remain stable across a wide temperature range (0.1 to 2.5).
- Optional Thinking Activation: Users can explicitly activate deeper thinking with "think deeply: prompt" or by using specialized Jinja templates for always-on thinking.
Performance & Benchmarks
The model demonstrates improved benchmark scores compared to its non-thinking predecessor, Gemma-3-27b-it-heretic, across various tasks like arc_challenge, arc_easy, boolq, hellaswag, piqa, and winogrande.
Decensoring Statistics
With a KL divergence of 0.07 (compared to 0 for the original model) and only 9 refusals out of 100, this model maintains high fidelity to the original while significantly reducing content refusals.
Recommended Use Cases
This model is suitable for applications requiring:
- Direct and Unfiltered Responses: Ideal for use cases where content filtering or refusal is undesirable.
- Enhanced Reasoning: Beneficial for tasks demanding complex problem-solving, detailed analysis, or improved understanding of visual inputs.
- Creative and Roleplay Scenarios: Its uncensored nature and reasoning capabilities make it well-suited for generating diverse and specific narrative content.