File	Size	Use Case	Quality
omega-v21-F16.gguf	~15 GB	Maximum quality, reference	Perfect
omega-v21-Q8_0.gguf	~8.0 GB	Highest practical quality	Excellent
omega-v21-Q6_K.gguf	~6.2 GB	Balanced (recommended)	Very Good
omega-v21-Q5_K_M.gguf	~5.7 GB	Lighter, good quality	Good
omega-v21-Q4_K_M.gguf	~5.3 GB	Smallest, mobile-friendly	Acceptable

Quick Start on Termux (Android)

bash

pkg install llama-cpp
wget https://huggingface.co/Abdllahd/OMEGA-V21-Full-Merged-GGUF/resolve/main/omega-v21-Q6_K.gguf

llama-cli \
  -m omega-v21-Q6_K.gguf \
  --ctx-size 4096 \
  --threads $(nproc) \
  -cnv \
  --chat-template gemma \
  --temp 0.3

Features

0% refusal rate (OBLITERATED base)
Bilingual: Arabic (70%) + English (30%)
ReAct reasoning with 6-point think structure
Real bash code (no placeholders)
Termux/PRoot-aware
Runs on 7.6GB RAM Android devices (Q4/Q5)

Hardware Requirements

Quant	Min RAM	Speed on Android
F16	18 GB	Reference only
Q8_0	10 GB	~2 t/s
Q6_K	8 GB	~3 t/s
Q5_K_M	7 GB	~3-4 t/s
Q4_K_M	6 GB	~4 t/s

Part of OMEGA v22 Architecture

Four-layer autonomous agent:

Spiders (data collectors)
Tools (action executors)
Orchestrator (this model)
SpatialCache (memory system)

Training Details

Base: OBLITERATUS/gemma-4-E4B-it-OBLITERATED (0% refusal)
Method: LoRA fine-tuning (rank 64)
Dataset: 10,000 bilingual examples (Arabic/English)
Format: Strict JSON with think reasoning + bash code blocks

Related Repositories

LoRA only: https://huggingface.co/Abdllahd/OMEGA-V21-Gemma4-LoRA-GGUF
Base backup: https://huggingface.co/Abdllahd/OMEGA-Gemma4-E4B-OBLITERATED-Base-Backup
Full merged (this): https://huggingface.co/Abdllahd/OMEGA-V21-Full-Merged-GGUF

License

Apache 2.0 (inherited from OBLITERATED base)