language:
- en
license: gemma
base_model: DavidAU/gemma-4-E4B-it-The-DECKARD-HERETIC-UNCENSORED-Thinking
tags: - gguf
- llama-cpp
- gemma4
- uncensored
- reasoning
- thinking
- heretic
- deckard
- roleplay
pipeline_tag: text-generation
Gemma 4 E4B Instruct - The DECKARD HERETIC UNCENSORED Thinking (GGUF)
This repository contains GGUF format quantizations for DavidAU's Gemma-4-E4B-it-The-DECKARD-HERETIC-UNCENSORED-Thinking.
These files allow for highly efficient local inference on CPUs and Apple Silicon, as well as VRAM-constrained GPUs, utilizing llama.cpp and compatible frontends.
🧠 About the Base Model
The base model is a specialized, heavily fine-tuned version of Google's Gemma 4 E4B (Effective 4B parameters).
- HERETIC / DECKARD Fine-Tune: This model has undergone the "HERETIC" abliteration and fine-tuning process by DavidAU. It is completely uncensored, unrestricted, and heavily calibrated for complex creative writing, roleplay, and deep fiction generation.
- Reasoning & Thinking: It includes augmented reasoning/thinking capabilities, allowing the model to internally chain logic before outputting text, making its responses deeper and more contextually accurate.
- Context Window: The Gemma 4 architecture supports extended context windows (up to 256k in base), making it highly capable for long-form conversations.
📦 Available Quantization Formats
This repository provides multiple levels of quantization to help you balance VRAM/RAM usage, generation speed, and model fidelity.
| File Name | Bit Resolution | Recommended Use |
|---|---|---|
| Q3_K_M | 3-bit | Ultra-low RAM usage. Noticeable perplexity degradation but runs on very constrained hardware. |
| Q4_K_M | 4-bit | Recommended. The sweet spot for local LLMs. Great balance of speed, low memory footprint, and quality. |
| Q5_K_M | 5-bit | Higher precision. Use this if you have the memory to spare and want slightly better reasoning. |
| Q6_K | 6-bit | Very high fidelity. Close to unquantized performance with a larger memory footprint. |
| Q8_0 | 8-bit | Extremely close to FP16 baseline. Best for users who want maximum precision and have ample RAM/VRAM. |