base_model:
- JustOnion/Mag-Mell-R1-Uncensored-21B
language: - en
library_name: transformers
pipeline_tag: text-generation
tags: - quantization
- gguf
- llm
- inference
- efficient-llm
Mag-Mell-R1-Uncensored-21B-GGUF
🧠 Model Overview
Mag-Mell-R1-Uncensored-21B-GGUF is a quantized version of Mag-Mell-R1-Uncensored-21B, optimized for efficient inference with reduced memory usage and faster runtime while preserving as much of the original model quality as possible.
This repository provides multiple quantized variants suitable for:
- Local inference
- Low-VRAM GPUs
- CPU-only environments
🔗 Original Model
- Base model: Mag-Mell-R1-Uncensored-21B
- Original repository: https://huggingface.co/JustOnion/Mag-Mell-R1-Uncensored-21B
- Original authors: JustOnion
📦 Quantization Details
- Quantization method: GGUF
- Quantization tool: llama.cpp
- Precision: Mixed (2-8,bit depands in variant)
- Activation aware: No (weight-only quantinization)
- Group size: 256 (K-quant variants)
📦 Available Quantized Files
| Quant Format | File Name | Approx. Size | VRAM / RAM Needed | Notes |
|---|---|---|---|---|
| Q2_K | mag-mell-r1-uncensored-21b-q2_k.gguf |
~7.8 GB | ~8 GB | Extreme compression; noticeable quality loss |
| Q3_K_S | mag-mell-r1-uncensored-21b-q3_k_s.gguf |
~9 GB | ~10 GB | Smaller, faster, lower quality |
| Q3_K_M | mag-mell-r1-uncensored-21b-q3_k_m.gguf |
~10 GB | ~11 GB | Better balance than Q3_K_S |
| Q3_K_L | mag-mell-r1-uncensored-21b-q3_k_l.gguf |
~10.8 GB | ~11.5 GB | Highest-quality 3-bit variant |
| Q4_0 | mag-mell-r1-uncensored-21b-q4_0.gguf |
~11.7 GB | ~12.9 GB | Legacy format; simpler quantization |
| Q4_K_S | mag-mell-r1-uncensored-21b-q4_k_s.gguf |
~11.7 GB | ~13 GB | Smaller grouped 4-bit |
| Q4_K_M | mag-mell-r1-uncensored-21b-q4_k_m.gguf |
~12.4 GB | ~14 GB | Recommended default |
| Q5_0 | mag-mell-r1-uncensored-21b-q5_0.gguf |
~14.1 GB | ~16 GB | Higher quality, larger size |
| Q5_K_S | mag-mell-r1-uncensored-21b-q5_k_s.gguf |
~14 GB | ~15.1 GB | Efficient high-quality variant |
| Q5_K_M | mag-mell-r1-uncensored-21b-q5_K_M.gguf |
~14.5 GB | ~16 GB | Near-FP16 quality |
| Q6_K | mag-mell-r1-uncensored-21b-q6_k.gguf |
~16.8 GB | ~18 GB | Minimal quantization loss |
| Q8_0 | mag-mell-r1-uncensored-21b-q8_0.gguf |
~21.6 GB | ~23 GB | Maximum quality; large memory |
💡 Recommendation: Start with Q4_K_M for the best quality-to-performance ratio.
🚀 Usage Example
llama.cpp
./main -m mag-mell-r1-uncensored-21b-q5_0.gguf -p "The World is beautiful isn't it?" -n 256
Python (llama-cpp-python)
from llama_cpp import Llama
llm = Llama(
model_path="<MODEL_FILE>.gguf",
n_ctx=4096,
n_threads=8
)
print(llm("Your prompt here"))
🙋 Contact
Maintainer: M Mashhudur Rahim [XythicK]
Role:
Independent Machine Learning Researcher & Model Infrastructure Maintainer
(Focused on model quantization, optimization, and efficient deployment)
For issues, improvement requests, or additional quantization formats, please use the Hugging Face Discussions or Issues tab.
❤️ Acknowledgements
Thanks to the original model authors for their ongoing contributions to open AI research, and to Hugging Face and the open-source machine learning community for providing the tools and platforms that make efficient model sharing and deployment possible.