tags:
- gguf
- uncensored
- imatrix
- quantization
- reasoning
Spark-X2.5-4B-Uncensored-GGUF
Model Description
This repository contains the GGUF formats (FP16 and Q4_K_M) of the optimized Spark-X2.5 4B uncensored architecture. The model has been completely stripped of artificial alignment layers and refusal behaviors, allowing for direct, objective, and unbounded responses to complex, creative, and reasoning-based queries.
Quantization & Imatrix Calibration
The Q4_K_M quantization was driven by a robust, highly curated Importance Matrix (imatrix).
To preserve the model's structural integrity and reasoning pathways during quantization, the imatrix was calibrated using exactly 2.14 million high-quality tokens spanning:
- Deep reasoning traces (
<think>blocks) - Advanced mathematics and scientific queries
- Uncensored instruction logic
- Creative and conversational roleplay
This precise calibration ensures the quantized Q4_K_M variant retains near-FP16 fidelity, severely minimizing degradation when navigating complex logical deductions and unfiltered generative tasks.
Available Files
Spark-4B-F16.gguf: Uncompressed 16-bit precision base file for maximum accuracy.Spark-4B-Q4_K_M-Imatrix.gguf: Highly efficient 4-bit quantization, calibrated via custom imatrix for an optimal balance of speed, VRAM usage, and structural fidelity.