Apertus

There would be a way under Australian law right now for an Australian startup to work with the government to license material through the existing regimes and do a deal under a license to get enough data, trading data to do the work.

Alternatively, you could take the Swiss approach like Apertus did and look for online data sources that are declared and uncopyrighted.

Another question this side.

Great discussion.

Quen 3, 3-2 base.

Apertus.

Arcee.

Barmba.

But a few countries are already treating AI the way Japan treated trains, as something the public pays for, so the public should own it.

Switzerland's public universities released Apertus, which is Latin for open, in 2025, a fully open model built with public money that anyone can use.

Spain, the same country that opened its high-speed rail to competition, released Alia, a government-funded open model built as its national supercomputing center.

The Netherlands is building GPTNL through a public research institute trained only on data it has the rights to use.

So we probably won't ourselves build our own frontier model or open model, but But we're in conversations with the people that do them.

So we talk with the Apertus team all the time here in Europe.

But I think about the entirety of the stack.

So if we think about on the application layer, the team behind Thunderbird, our email client, has built something they call Thunderbolt, which is basically their... their chat client, which has access to your context, has access to your calendar, has access to your history.

And then the smaller sovereign initiatives, which is where the count is really rising.

Switzerland's Apertus from ETH Zurich and EPFL released September twenty twenty-five.

The seventy billion version was trained on the Alps supercomputer, and the training data and methods are fully disclosed, which is unusual.

Fully disclosed meaning you could rerun it.