Luna

How it's tested

Built and tested at home.

Luna is only as capable as the models it runs on. These are its defaults: what it runs every day, and what every change is tested against. Swap any of them for your own.

The default setup.

The model it thinks with

An open-weight 9B model that's good at tool calling, run at 4-bit on llama.cpp. Small enough for a laptop: it's the size Luna is built for.

Everything else, on the same machine

A small model for finding things by meaning, a smaller one for background jobs like writing up a conversation, and speech, voices and pictures, all running locally.

What it runs on

Tested on Apple silicon Macs with 16 to 32 GB of memory. Where it's heading: a Mac mini at home, or a computer of your own elsewhere, always on for the whole family.

How it's tested.

01

Everyday requests

A set of the kinds of thing people actually ask, from “what's on tomorrow?” to “cover me for an hour”, each checked for whether it does the right thing.

02

Requests it's never tuned on

Separate sets kept back on purpose, so it isn't only good at the examples it was shaped around.

03

Nothing gets worse

Changes are measured against both before they go in. Anything that makes a request that used to work stop working is left out.

Good to know.

Can I use a bigger model?

Yes. Any model your machine can run that handles tools will do. A bigger one on a bigger machine does better; a smaller one does less.

Will the defaults change?

As better open models come out, yes. Whatever the default is, it's what everything is tested against.

Does it need the internet?

Not to think. The models run on your machine; only things that are online by nature, like the web or email, need a connection.

Coming soon

It isn't out yet. If you'd like to hear when it is, get in touch.

Get in touch