Last time we told you what changed and why. This time we want to introduce the thing itself.
vinci-studio ships August 8: a capable model built on a strong open base, shaped through our Constitutional Fine-Tuning process, and shipped open under Apache 2.0. It's the first of a family we're calling Vinci.
Here's what it is, where the name comes from, and — the part we care about most — how you can check that it does what we say.
The name
Vinci is for Leonardo, but not for the reason you'd guess. It's not about genius.
It's about a habit of his he called saper vedere — "knowing how to see." The idea that understanding something means looking at it closely and honestly, from every side, before you decide what's true.
That's the posture we want for a model used in serious work. Not a confident voice that sounds right. Something that looks at what's actually in front of it, shows you what it found, and lets you see the same thing it saw.
The family grows from there — bozza, studio, tela, opera, magna, codice — each one a different tool for a different kind of work. studio is first. We'll walk through the whole lineup in a few weeks.
The thing we actually want you to know
We're publishing everything. The methodology. The constitution the model is trained against. The character document. The adversarial test results from the exact weights we deploy.
Not because we're required to. Because we don't think the thing that earns your trust should be a secret.
A lot of AI companies ask you to trust a method you can't see and a model you can't inspect. We're doing the opposite. The recipe is public, the spec is public, and the test results are public — so the only thing left for us to be judged on is whether Vinci actually behaves the way our constitution says it will.
That's the point of all of it. Not "trust us, we have a clever method." Closer to: here's exactly what we promised, here's how the model actually behaved, line them up yourself.
If they don't match, you'll know before we say a word. That's by design.
What defines Vinci
Five things, and each one is something you can verify rather than something you have to believe.
Capable. It stands on a frontier-grade open base, so it's genuinely smart from the first prompt. We didn't rebuild the engine — we built everything around it.
Character. We train it against a written constitution and a character document. Both are public and both ship with the weights, so you can read exactly what Vinci is meant to do and who it's meant to be before you ever deploy it.
Verifiable. We publish real adversarial test results — HarmBench, JailbreakBench, and a testbed for Chinese-model censorship behavior, run against the model we actually ship, not a cleaned-up demo. Then you run your own.
Grounded. It answers from your documents and cites them. No source, no claim. If it can't point to where an answer came from, it shouldn't be giving you the answer.
Sovereign. Hosted on Canadian sovereign infrastructure in Quebec, with zero data retention by default. Your data stays in the country and doesn't stick around.
The promise behind it
A specification only means something if someone's on the hook when reality drifts from it. So here's ours, plainly.
If Vinci's deployed behavior ever diverges from what the verification bundle says, we publish the diagnosis and the fix within 30 days.
We can make that promise because the weights and the tests are open — you can catch the drift yourself, which means we can't quietly pretend it didn't happen. Closed providers can't offer the same thing, because their customers have no way to confirm the divergence in the first place.
We'd rather be held to a promise you can check than be trusted on one you can't.
What we're not claiming
Honesty cuts both ways, so a couple of things we want to be upfront about.
This whole approach is only as strong as the tests are hard. If our adversarial tests were weak or we graded our own homework, "verifiable" would just be a nicer-sounding kind of marketing. So we're building the tests to be genuinely adversarial and runnable by anyone — and we'd rather you stress them than trust them.
And as we said last time, the dedicated Quebec legal-French work is waiting for now. Vinci is strong in French as a capable, multilingual model, and French sources live in the library it reads from — but the polished Quebec legal register comes later, not at launch.
What's next
Next week we get into how the Constitutional Fine-Tuning actually works, in plain language — the part everyone assumes is the secret, written out so you can use it yourself.
After that: the open release of our legal corpus and evaluations, the thinking behind the 30-day promise, the full Vinci family, and Vinci's character — landing on launch day.
New post every Tuesday at 9am ET. The next one's July 7.
vinci-studio ships August 8. When it does: read the constitution, fork the weights, run the tests. See if it does what we said.
Talk soon, The SimpleDirect team
Where to next