Airunner: Offline Multi-Modal AI Engine for Local Inference

Project Overview

Airunner is an open-source offline AI inference engine that lets you run image generation, real-time voice conversations, LLM chatbots, and automated workflows entirely on your local machine. No internet required, no data leaves your computer. Perfect for privacy-conscious users and developers who want full control over their AI tools without cloud dependencies.

If you've been poking around the local AI scene, you might have stumbled upon a tool called Airunner. Maintained by Capsize-Games, this open-source project is essentially an offline inference engine — you can generate images, have voice conversations, chat with LLMs, and even chain multiple tasks into automated pipelines. All of it runs on your own hardware, no cloud subscriptions required. The catch? You'll want a decent GPU to make it sing.

What Can It Actually Do?

At its core, Airunner provides a modular pipeline system. It bundles different AI models into a unified interface, letting you compose workflows using visual nodes. The tool covers four main areas:

Image generation: Supports Stable Diffusion models, including text-to-image, image-to-image, ControlNet, and LoRA adapters.
Real-time voice chat: Integrates Whisper for speech recognition and various TTS engines like Tacotron and Coqui for low-latency voice interaction.
LLM chatbot: Can load popular open-source models like Llama, Mistral, and Gemma for local conversation.
Automated workflows: Connect the above modules together — think “voice input → LLM processing → voice output” or “image generation → LLM description → log to file”.

Since everything stays on your machine, your conversations, generated images, and audio files never leave your hard drive.

Who Actually Needs This

Airunner shines if you're privacy-focused or often work offline. Consider a freelance illustrator who wants to generate concept sketches without uploading client work to a third-party server. Or a developer testing a chatbot before deploying to production — you can iterate quickly in a local environment. Hobbyists building a voice assistant will appreciate the nearly turnkey real-time speech pipeline.

But let's be clear: it's not as simple as a web-based one-click tool. You'll need a Python environment and ideally a GPU with at least 8GB VRAM. CPU inference is possible but noticeably slower, especially for voice conversations.

After spending some time with it, my take is that Airunner feels more like a toolbox than a polished consumer app. It ships with a few preset workflows, but the real power comes from building your own node arrangements. If you're comfortable with node editors like ComfyUI or Blender's shader nodes, you'll feel right at home. For everyone else, the learning curve might be a bit steep.

Getting Started and Caveats

Installation is straightforward — there's a pip package and a one-click script for Windows. After launch, you're greeted by a node editor with a model library on the left and a canvas in the middle. You'll need to manually download model weights from Hugging Face and point Airunner to the right folders. Once running, performance depends heavily on your GPU: an RTX 3060 handles TinyLlama chat nearly instantly, and SDXL takes about 20 seconds per image.

The voice module impressed me. I said “Hello” into a microphone, Whisper transcribed it locally, the output went to an LLM, and the reply was spoken via Coqui TTS — total latency under 3 seconds. By swapping to smaller models like distil-whisper and XTTS-v2, you can cut latency further.

Downsides? The codebase and community documentation are a bit sparse; you'll end up digging through GitHub Issues or the Discord server for advanced usage. Additionally, loading both image and voice models simultaneously can strain VRAM — 8GB is barely enough for SDXL plus Llama 7B without choking.

Practical Advice

Start small. Try TinyLlama and SD 1.5 first to verify your setup, then move up to 7B+ models. Take advantage of the automation workflows — linking image generation to an LLM-based description can save hours of manual annotation.

Overall, Airunner stands out as one of the more comprehensive offline AI engines available right now. It's ideal for anyone who wants total data control and is willing to invest time in configuration. If you just need a simple chat window or a straightforward image generator, tools like Ollama or Stable Diffusion WebUI might be easier. But Airunner gives you the power to tie them together — and that flexibility is genuinely valuable.