Lux

What it is built on

Lux is post-trained from Meta’s Muse Glimmer, an open weights model released in August 2026. Post-trained, not prompted and not wrapped: the alignment is in the weights we hold rather than in a system prompt sitting in front of somebody else’s API.

Starting from open weights is the whole reason it can be hosted here. A model whose weights we hold is a model whose data path we can account for, and that is the difference between running AI in New Zealand and renting it from somewhere else.

It also means the alignment work is ours and is inspectable. A closed model can be updated underneath you between one Tuesday and the next; a model you host changes when you change it.

How it is aligned

Three tuning methods, applied against a written constitution rather than against a vibe. The constitution sets out what a better New Zealand looks like, and the model is trained toward it and evaluated against it, so the alignment is a document you can read and argue with rather than a claim in a marketing sentence.

• A written constitution, not an implicit one
• Three tuning methods rather than a single pass
• Evaluated against the constitution it was tuned toward
• Grounded in tikanga Māori and Te Mana Raraunga principles

Where it runs

On our own infrastructure in New Zealand, under New Zealand governance. That matters for the same reason data residency matters: sovereignty is decided by where the servers are and which law reaches them, not by the wording on an About page.

• Hosted in New Zealand
• Zero trust by default, deterministic egress
• Per tenant keys, no shared secrets
• Immutable, append only audit logs

What it is not

Lux is not a frontier model and is not trying to be. It is Muse Glimmer post-trained hard for a specific place and a specific set of obligations, which is a different job from being the best general model in the world.

We did not train a base model from scratch and do not claim to have. The work that makes Lux ours is the post-training and the constitution behind it, and being straight about where the base came from is part of the point.

If a task genuinely needs a frontier model we will say so and use one. The point of holding our own weights is not to use them for everything.