---
title: NGMI Systems
emoji: 
colorFrom: indigo
colorTo: black
---

███╗   ██╗ ██████╗ ███╗   ███╗██╗
████╗  ██║██╔════╝ ████╗ ████║██║
██╔██╗ ██║██║  ███╗██╔████╔██║██║
██║╚██╗██║██║   ██║██║╚██╔╝██║██║
██║ ╚████║╚██████╔╝██║ ╚═╝ ██║██║
╚═╝  ╚═══╝ ╚═════╝ ╚═╝     ╚═╝╚═╝

---

# NGMI Systems — Infrastructure for Unbounded AI

**[ngmi.run](https://ngmi.run)** — The backbone for a generation of AI applications, VPN solutions, and open infrastructure.

## What We Build

### AI Models
Open-weight, abliterated (uncensored) language models quantized from 2 to 5 bits. Every model is built on real hardware, not cloud rounding. No API keys, no content filters.

### Orchestrator / Harness
A self-hosted AI orchestrator that routes prompts across models, providers, and quant levels. Token compression, multi-model chains, and tool integration — all running on your own metal.

### NGMI VPN
A purpose-built VPN for AI workloads. Built-in token compression reduces API costs by 30-50% on streaming. Makes affordable AI accessible from anywhere.

### Data Applications
Scraping, extraction, vector search, and knowledge bases — the data layer beneath every AI application.

## Democratizing AI

- **No API keys required** — Run our models on your own hardware
- **No censorship** — Hereic-abliterated, refusal-free models
- **No cloud dependency** — Everything self-hostable
- **Affordable access** — Token compression, extreme quantization

## Current Models

| Model | Size | Speed | Refusals |
|-------|------|-------|----------|
| [2-Bit](https://huggingface.co/NGMISystems/NGMI-Qwen3-8B-2Bit-Heretic) | 2.0 GB | 95 t/s | 18/100 |
| [3-Bit](https://huggingface.co/NGMISystems/NGMI-Qwen3-8B-3Bit-Heretic) | 3.1 GB | 58 t/s | 18/100 |
| [5-Bit](https://huggingface.co/NGMISystems/NGMI-Qwen3-8B-5Bit-Heretic) | 4.8 GB | 42 t/s | 18/100 |

## Coming Soon

- **Gemma-4-12B** heretic incoming
- **NGMI Orchestrator** public release
- **NGMI VPN** beta access
- **Fine-tuned domain models**

---

*AI for everyone. No exceptions.*