News

Same team, same open-source Matricxon and pAIring, a new name and address: bytepenny.com. Old upupabit.ai links keep working and lead here.

Three more cloud connectors: 100+ models from NVIDIA with free developer credits, Cohere with a free trial key and embeddings, and your own RunPod endpoint. The connectors page now compares all eight.

Three more cloud connectors, all free to try: hundreds of models through OpenRouter, Google Gemini with embeddings for the knowledge base, and fast open models on Groq.

A second cloud connector: dozens of open and commercial models, plus embeddings for the knowledge base, with one access key.

Run chats on dozens of cloud models, from Gemma 4 to GPT-OSS, plus embeddings for the knowledge base. Nothing to deploy.

A new home for Matricxon and pAIring, with full documentation for both.

Several Matricxon nodes can now share models and chat caches, managed from a live node map in pAIring.

Connect MCP tool servers and let models look things up or take actions in chat.

Matricxon keeps long chats ready on disk with AES-256-GCM, so returning to them is fast.

Matricxon can run on an NVIDIA GPU, with packed or unpacked weights.

Matricxon runs Qwen 3.5, including its hybrid layers and vision.

Matricxon runs Gemma 4, and mixture-of-experts models share one faster expert layer.

Three more model families join Matricxon, including its first mixture-of-experts model.

Matricxon reads 25 quantization types, now including the IQ1 to IQ4 types for very small models.

pAIring shows math formulas in replies as clean, readable equations.

pAIring can send chats to engines on cloud GPU services when you need a bigger model.

Matricxon moved all dequantization to native C kernels, many times faster than before, with macOS support the same day.

The first public release of Matricxon, a readable GGUF runtime with Qwen 2 and Qwen 3 support.

pAIring 1.0, a private AI chat workspace for teams, is released under AGPL-3.0.