Trained on Dawn.

Our training allocation on Dawn — the UK's fastest AI supercomputer, at the University of Cambridge — is complete. This note records what we did there, what it means for the people we build for, and what we are deliberately not saying.

What happened

Through the UK's AI Research Resource (AIRR), Aevum trained the shared base models behind our product family: the teaching assistant, school operations, marking and feedback, tutoring, and multimedia. Alongside the models themselves, we built the evaluation evidence around them — safety testing, bias auditing, and red-teaming took more of our allocation than any feature did. That ratio was deliberate. All products remain in development.

Why it matters to a school

Everything was trained on British national infrastructure, under UK law, on licence-clean data — exam-board public documents, open educational resources, and synthetic material we generated ourselves. No customer data was involved, because we had none to involve and our architecture never requires it. The next stage brings the models home: to the schools and organisations that will one day own them, with the data staying in the building. Own the model. Keep the data. Inspect the reasoning.

What we're not saying, and why

You will notice this note contains no benchmark scores, no training recipes, and no technical specifics. The scores are unpublished until our pre-registered evaluation protocols are complete — we publish measurements, not marketing. The recipes are our intellectual property. A company asking schools to trust it should be exact about what it shares and honest about what it keeps; this page is both.