| --- |
| title: NGMI Systems |
| emoji: |
| colorFrom: indigo |
| colorTo: black |
| sdk: static |
| pinned: false |
| --- |
| |
| # 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, no one telling you what you can ask. |
|
|
| ### 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. No vendor lock-in. |
|
|
| ### NGMI VPN |
| A purpose-built VPN service designed for AI workloads. Built-in token compression reduces API costs by 30-50% on streaming responses. Routes traffic through optimized paths while preserving privacy. Makes affordable AI accessible from anywhere. |
|
|
| ### Data Applications |
| Scraping, extraction, vector search, and knowledge bases — the data layer beneath every AI application. Real-time Reddit ingestion, ClickHouse analytics, medical imaging pipelines. |
|
|
| ## Democratizing AI |
|
|
| The AI industry is consolidating behind paywalls, content filters, and corporate APIs. We build the alternative: |
|
|
| - **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, optimized inference |
|
|
| ## Our Stack |
|
|
| | Layer | Technology | |
| |-------|-----------| |
| | Models | Qwen, Gemma, PrismML ternary — abliterated + quantized | |
| | Inference | llama.cpp CUDA, ONNX Runtime | |
| | Orchestration | C++/Java/C# harness, WebSocket streaming | |
| | Data | ClickHouse, Parquet, custom scrapers | |
| | Infrastructure | NVIDIA DGX (GB10), Tailscale mesh, Contabo VPS | |
| | VPN | Custom WG-based, token compression layer | |
|
|
| ## Current Models |
|
|
| - [NGMI-Qwen3-8B-2Bit-Heretic](https://huggingface.co/NGMISystems/NGMI-Qwen3-8B-2Bit-Heretic) — 2.0GB, 95 tok/s |
| - [NGMI-Qwen3-8B-3Bit-Heretic](https://huggingface.co/NGMISystems/NGMI-Qwen3-8B-3Bit-Heretic) — 3.1GB, 58 tok/s |
| - [NGMI-Qwen3-8B-5Bit-Heretic](https://huggingface.co/NGMISystems/NGMI-Qwen3-8B-5Bit-Heretic) — 4.8GB, 42 tok/s |
|
|
| ## Coming Soon |
|
|
| - **Gemma-4-12B** heretified |
| - **NGMI Orchestrator** public release |
| - **NGMI VPN** beta access |
| - **Fine-tuned domain models** (code, medical, creative) |
|
|
| --- |
|
|
| *AI for everyone. No exceptions.* |
|
|
