Pulsar AI Research

Australian-made AI, trained only on Australian text.

SpaceAI is our core research line. Every layer is built from scratch: the corpus, the tokenizer, the model, and the system that trains it across ordinary GPUs. No borrowed model weights and no overseas training framework.

Download the corpus SpaceAI on GitHub

What exists today

SpaceAI Australian Corpus v1
53,337 documents, about 484 million tokens. Legislation, Productivity Commission, AIHW, RBA and Treasury publications, a century of Year Book Australia, and public-domain Australian literature. CC BY 4.0, with attribution kept on every document.
Released
spaceai-sp32k tokenizer
A 32,768-token vocabulary trained only on Australian text, so names like Wagga Wagga and words like Commonwealth stay whole.
In use
SpaceAI-0.1B
A 100.7 million parameter model trained from random initialisation on the corpus. Weights are coming for Australian users only, with each access request approved individually.
Training
Swag
Our system for training one model across the everyday GPUs Australians already own, instead of renting a data centre. It is ours and it is not published.
In use

What goes into an Australian corpus

  • Australian-origin material only, and only what is free to redistribute: CC BY 4.0 or public domain.
  • No court decisions, Hansard, AustLII or share-alike material.
  • No web text after November 2022, and no AI-generated text.
  • Third-party material that publishers exclude from their licences is withheld.
  • Personal contact details are redacted, and every document keeps its attribution.

Who can use what

Corpus

CC BY 4.0. Free for anyone, with attribution. The public-domain books are free of copyright in Australia and the US.

Get the corpus

Model

SpaceAI Community Licence. Australian residents and organisations, using it in Australia. Free, including commercial use.

Read the licence

Training system

Swag, the training code and the tokenizer pipeline are proprietary and not published.

Working on Australian AI?

Research collaboration, compute partnerships, or early interest in SpaceAI models for your organisation: send us a note.

Get in touchSee the roadmap