What we've shipped, and what comes next.

Each track is listed in order, without dates. A step moves up when the work behind it is done and the evidence is ready to show.

SpaceAI models and corpus

Small models trained from scratch on Australian text, released to Australians first.

Now Australian Corpus v1 published: 53,337 documents, about 484 million tokens. Released
Now SpaceAI-0.1B training from random initialisation on the corpus, with the sp32k tokenizer. Training
Next SpaceAI-0.1B weights for Australian users under the SpaceAI Community Licence, with gated download. Planned
Then Corpus v1.1: more Year Book Australia issues and more public-domain Australian literature. Planned
Then Larger SpaceAI models. Planned
Later SpaceAI offerings for Australian organisations. Planned

Training one model across the everyday GPUs people already own.

A local-first workspace where models, machines and human approval share one loop.

Now Private beta of PulSen Craft Desktop and Fleet. Access is closed while the IP is secured. Private beta
Next Scoped organisation pilots, each with its own workflow, data boundary and evaluation plan. Planned
Next A SpaceAI model replaces the stand-in starter model. Planned
Later Self-improving agents and a training feature. Not in the current beta. Planned

Claims about models and products should be checkable by someone else.

Now Five Pillars of AI Safety: 21 open-weight models, five behaviour tests, published July 2026. Published
Next A capability register: claim, implementation, evidence and demonstration method for each major feature. Planned
Next Reproducible experiments with methods, data, results, limitations and safe reproduction notes. Planned
Then A dated provenance timeline backed by primary artefacts. Planned

We're looking for research collaborators, compute partners and organisations willing to scope a pilot.