---
license: apache-2.0
language:
  - en
tags:
  - ngari
  - sovereign-ai
  - edge-ai
  - qwen
  - gguf
  - ollama
  - teacher-model
  - distillation
---

# NGARi Sovereign Edge — Qwen 27B Serving Configuration

This repository contains the **serving configuration and model card** for running a Qwen 27B-class instruct model (Qwen3.5 family, 27.3B params, Q4_K_M GGUF) on **NGARi Orin 64GB edge hardware** via Ollama — with zero cloud dependency and a verified air-gap.

> This repo ships configuration (Modelfile + card), not weights. Weights are pulled through Ollama's registry under the tag `qwen3.8:27b`.

## The role of this model in the NGARi stack

The 27B model is NGARi's **R&D engine**, not just an end product. It powers the creation of NGARi's small, fast edge models:

| Use | What the 27B does | Output |
|-----|-------------------|--------|
| **Knowledge distillation** | Generates reasoning traces / logits as a teacher | Smarter 1.5B students (e.g. [ngari-ft-distilled](https://huggingface.co/NGARiAI/ngari-ft-distilled)) |
| **Synthetic data** | Creates diverse, high-fidelity training examples | [ngari-datasets](https://huggingface.co/NGARiAI/ngari-datasets) |
| **Tool-use training** | Simulates tool schemas, queries, correct calls | [ngari-tool](https://huggingface.co/NGARiAI/ngari-tool) (100% tool-format eval) |
| **LLM-as-a-judge** | Scores student outputs for DPO/preference tuning | Aligned small models, fewer hallucinations |
| **Sovereign specialization** | Masters private/local data in a secure environment, then distills | Custom edge models that preserve privacy |

The result is the "holy trinity" of edge deployment: **accurate** (teacher distillation), **powerful** (specialized tool use), **fast** (small parameter count).

## Key features

- Runs on NVIDIA Jetson AGX Orin 64GB (aarch64) — Q4_K_M, ~16GB footprint
- 262K-token context window; completion, tools, thinking, vision capabilities
- Zero cloud dependency for inference; verified air-gap (monitored via `/proc/net/dev`)
- Apache 2.0 kernel: [github.com/NGARiAI/ns-bos-kernel](https://github.com/NGARiAI/ns-bos-kernel)

## Usage (Ollama)

```bash
ollama run qwen3.8:27b "your prompt here"
# long-lived sessions:
ollama run qwen3.8:27b --keepalive 20m "your prompt"
```

## Licensing

- Qwen base model: Apache 2.0 (Qwen license terms apply to the weights)
- NGARi-authored configuration and pipeline artifacts: Apache 2.0 — Bison Venture Partners LLC dba NGARI

## Sovereign AI

All inference runs on user-owned edge hardware. Zero cloud dependency. Verified air-gap. *"AI You Own. Completely."*