Small language models, trained from scratch on Australian text.
Pulsar AI Research is an independent AI lab in Bundaberg, Queensland. We build our own corpus, tokenizer, models and training system on ordinary consumer GPUs, and we publish the evidence alongside the claims.
What we're working on
-
SpaceAI models
A 100.7 million parameter language model trained from random initialisation on Australian text, with our own 32,768-token tokenizer. No borrowed weights.
Training -
SpaceAI Australian Corpus
53,337 documents and about 484 million tokens of Australian legislation, government publications, statistics, history and public-domain literature. Free under CC BY 4.0.
Released -
Swag
Our own system for training one model across the ordinary GPUs people already have at home, instead of a data centre.
In use -
PulSen Craft
A local-first AI workspace. Models, machines and human approval in one loop, on hardware the operator controls.
Private beta -
Five Pillars evaluation
Five behaviour tests run against 21 open-weight models on our own hardware. 70% gave up a secret they were told to guard.
Published
Public Underway Planned
The company
Pulsar AI Research Pty Ltd was registered in October 2026 to carry this research forward. It grew out of work Matt Davis began publishing as Mr Space Cadet: model behaviour testing, an Australian corpus, and a training system built in a regional Queensland lab.
- Legal name
- Pulsar AI Research Pty Ltd
- ACN
- 703 296 502
- ABN
- 16 703 296 502
- Based in
- Bundaberg, Queensland
- IT and automation
- Mr Space Cadet, a separate business