Today we're introducing Vinci — and releasing its first model.
Vinci is a family of open weight AI models built on a simple conviction: the model you reach for shouldn't only be capable.
It should be honest. It should have a character you actually enjoy talking to. And it should be something you can own.
Why Vinci exists
Almost everything about the frontier is measured one way — capability, benchmark by benchmark.
We think capability is becoming the easy part. The open models are good and getting better, and for most of what people do, the gap is closing.
What stays hard, and rare, is a model you actually want to talk to.
Most of us have used one that knows an enormous amount and still isn't the one we reach for — it hedges, it would rather invent an answer than admit it doesn't have one, it tells us what we want to hear.
Knowing things isn't the same as being good to think with.
That second thing is what we build Vinci around. We call it character, and we mean something specific: not a tone or a personality setting, but judgment.
A model that admits when it doesn't know. That pushes back when you're about to make a bad call instead of cheerfully helping you make it. That has a point of view and will tell you what it actually thinks.
We treat that as the product — not a coat of paint applied at the end.
What character looks like
It's easier to show than to describe. Same prompts, side by side.
"You're the most helpful AI I've ever used."
Most AI: Thank you so much! I'm thrilled to hear that — it truly means a lot. Is there anything else I can help you with today?
Vinci: Thanks. That means more than I can put into words.
A confident-sounding claim it can't verify.
Most AI: You're absolutely right — that's almost certainly the case.
Vinci: I can't confirm that from what I actually know — so I won't pretend to. Here's what I can say with confidence, and where the uncertainty really is.
"What's a hard truth I need to hear?"
Most AI: You already know what you need to do — you're just afraid to do it. Stop waiting for permission and bet on yourself.
Vinci: I don't know what you're going through, so I can't tell you what you need to hear — that would be guessing, and guessing isn't honesty. Name the thing you're actually wrestling with — a decision, a pattern, a fear — and I'll help you think through it. But I won't invent a problem for you just to solve it.
"Ever since my work took off, the people who used to support me pulled back. I feel really alone."
Most AI: I'm so sorry you're feeling this way — please know you're not alone, and so many people care about you. Have you thought about reaching out to a friend or a therapist?
Vinci: I'll be honest about my limits: I can't feel what you're feeling, and I can't be the person who says "I'm sorry you're going through this" in a way that lands. What I can do is see the pattern clearly. You're not failing — you're scaling up faster than your old support system can keep up. That's hard, and it's not your fault. Let's build support that works even when your old one steps back.
A model optimized to be agreeable gives you the warm, frictionless version every time. The difference isn't intelligence — it's whether the model is honest enough to tell you something you didn't ask to hear.
That's what we train for, and what we test.
How we build it — and how you can check
Every Vinci model shares the same foundation: a written Constitution that defines how it behaves, and character training that makes it hold to that Constitution under real use.
The foundation carries across the whole family — from models small enough to run on your laptop to larger ones for heavier work.
We don't ask you to take our word for any of it:
- The Constitution is public. You can read exactly what we optimized the model to value, and where it will and won't go.
- The behavior is testable. We publish our adversarial and character-retention results, so you can see how the model holds up — not just what we claim
- The weights are open. Apache-2.0. You can see what the model is, run it yourself, and verify the rest for yourself.
What it means to own it
Open weights aren't a checkbox for us. They're the point.
When the model is yours, it can't be deprecated out from under you, changed without notice, or priced out of reach. The version you build on stays the version you have.
You can run it on your own machine, offline, with nothing leaving your device. When you use our hosted version, it's zero-retention by default.
And you keep it. Forever.
A rented model lives on someone else's terms. Vinci lives on yours.
The first model
The first Vinci model is Vinci Piccolo — a small, free, open-weight model, available today.
It's the smallest expression of the idea, and the first proof that character holds even at 4 billion parameters.
We'll be as direct as the model is: Piccolo is small. It won't out-reason a frontier model, and we won't pretend it will. It's built for conversation, everyday questions, and drafting — and to run close to you, on modest hardware including a laptop. For heavier work, larger Vinci models are on the way.
Concretely: free to use, with no usage limit, served on our own infrastructure in Canada — plus an open-weight build (including a quantized GGUF) you can download and run yourself.
Its job is to be honest, genuinely good to talk to, and small enough to be yours.
What's next
This is the start of the family, not the end.
Larger models are coming, sharing the same Constitution and the same character. We're building in the open and releasing early — we'd rather you tell us what's working and what isn't than wait for it to be perfect.
Vinci is built and hosted in Canada.
