About

Private AI you can read, run and own

Two open-source projects that bring modern AI to your own hardware, without giving up your data.

Our story

pAIring began as an AI student's research paper project: a simple, private chat workspace for local models. Running it with Ollama for real teams showed what was missing: control over memory, caching and many users at once.

So we started writing our own engine inside pAIring. That engine became Matricxon, an inference runtime written from scratch in Python and PyTorch, with optional native C kernels. Today the two are built side by side, and Matricxon is pAIring's main engine.

Our name comes from the hoopoe (Upupa), a gentle crested bird, and a few values travel with it. We believe in going vegan and being kind to animals. And we believe software should be open and free, as in beer.

projects, built together: 2
lines of llama.cpp or ggml: 0
quantization types: 25
pAIring API endpoints: 90+

What we believe

Private by default

Your models, chats and documents stay on machines you control. Your prompts are never used for training, and nothing is sent out unless you choose it.

Readable over clever

Plain Python and plain HTML, written so anyone can follow, check and change every step.

Use what you have

A laptop CPU should be enough to start. Bigger hardware and more nodes make it faster, not possible.

Free as free beer

Open source you can study, change and share, and free to download and run. No fees and no locked features.

Kind to animals

We are named after a bird, and we care about the creatures we share the world with. We believe in going vegan.

Honest about limits

BETA and experimental parts are marked clearly, and the docs say what is not tested yet.

Roadmap

Where the projects are heading. Plans change as we learn, and the full lists live in each repository.

Now In BETA, being hardened
- Matricxon cluster mode, with shared models and chat caches
- Cluster management and the live node map in pAIring
- Signed admin calls between pAIring and Matricxon
Next Planned
- Voter-only nodes on the Cluster page
- Smarter chat routing between cluster nodes
- Large cluster tests on GPU nodes
- Faster CPU decoding, closer to llama.cpp
Later Ideas we are exploring
- More model families, such as GLM-4 and DeepSeek MoE
- Vision for more architectures, starting with Gemma
- OCR for scanned PDFs and a real vector index
- Image generation on macOS

Open source

Our projects are released under GNU AGPL-3.0. Run, change and self-host them for personal or internal use. If you offer a changed version to others as a hosted service, share those changes back.