A quick and honest update, because parts of this site are currently telling you things we no longer believe.
Over the weekend of June 13–15 we changed our minds about something big. We're pulling our model work together under one name — Vinci — and the first one, vinci-studio, ships August 8.
We're writing this now instead of waiting for launch day because the old plan is still up on getsimpledirect.com, and we'd rather just tell you what's going on than let you read a version of us that doesn't exist anymore.
What changed
If you've been following along, you know the flash series — flash-1-mini, flash-1, flash-1-pro — as Canadian-specialized models we were building from scratch. We're moving on from that.
The short version of why: we were trying to build a model that had memorized Canada. What actually works better is a genuinely capable model that can look things up — pulling answers from a live, trusted set of sources the moment you ask, and showing you exactly where each answer came from.
Think open-book exam, not closed-book. The "Canadian" part lives in those sources, kept current, not baked into a model months ago and slowly going stale.
That one change rearranges everything else, and it came down to three things we couldn't ignore.
First, the open models you can download and run yourself got good. Really good. DeepSeek's V4 Flash is now strong enough that shaping it beats anything we'd build from zero — faster, cheaper, and we can still ship the result openly. Building our own engine from scratch stopped making sense the week someone handed us a better one for free.
Second, our customers had been telling us the real requirement the whole time. Banks, hospitals, government teams — across the board, what they need first is behavior they can verify, answers they can trace to a source, and hosting that stays in the country. Canadian legal depth matters to some of them. The trustworthy, capable, grounded part matters to all of them. We'd been leading with the wrong thing.
Third, the thing we're actually best at — taking a strong open model and shaping it into something well-behaved, consistent, and checkable — turns out to be useful far beyond Canada. So instead of one Canadian model, we're building one approach that serves Canada first and then anywhere with the same problem. Which is everywhere.
One honest tradeoff
Focusing on the best general model means our dedicated Quebec legal-French work has to wait for now.
We're not dropping French — and we want to be straight about how it comes back. A capable, characterful model is already fluent across the major languages, French included, and the French sources live right in the library it reads from. The polished Quebec legal register comes later rather than at launch. We'd rather say that plainly than pretend the roadmap didn't move.
What Vinci is
vinci-studio is built on DeepSeek V4 Flash, shaped through our Constitutional Fine-Tuning process, and shipped open under Apache 2.0. Five things define it:
- Capable — it stands on a frontier-grade open base, so it's smart from the first prompt.
- Character — we train it against a written constitution and a character document. Both are public, both ship with the weights, so you can see exactly what it's meant to do.
- Verifiable — we publish real adversarial test results (HarmBench, JailbreakBench, and a Chinese-censorship testbed) run against the model we actually deploy, not a cleaned-up demo.
- Grounded — it answers from your documents and cites them. No source, no claim.
- Sovereign — hosted on Canadian sovereign infrastructure in Quebec, with zero data retention by default.
The name is from Leonardo da Vinci — his idea of saper vedere, "knowing how to see." More on the name and the rest of the family next week.
What ships on August 8
- vinci-studio's open weights on Hugging Face, under Apache 2.0
- a live endpoint at try.getsimpledirect.com
- four things you can hold us to: the Constitution (v1.0), the Character document (v1.0), the methodology writeup (v2.0), and the adversarial test results from the deployed weights
The idea is simple. The model does what the bundle says, and you don't have to take our word for it — run your own tests against the open weights and check.
We're also making a promise we can actually keep: if Vinci's behavior ever drifts from what the verification bundle says, we publish the diagnosis and the fix within 30 days. Closed models can't offer that, because you'd have no way to confirm the drift in the first place. We can, because the weights and the tests are open.
What happens to the flash work
flash-1-mini stays up on Hugging Face as a research artifact. flash-1-pro is on hold while we focus on Vinci. None of it was wasted — it's the groundwork the methodology and the tests are built on. We'll keep you posted as things firm up.
And two things we're genuinely proud of are going out to everyone: the Canadian Bilingual Legal Corpus and CBLRE, our way of measuring how well an AI handles real legal questions. They were going to be private training fuel. Instead we're releasing them as open research anyone in the world can use, test, and build on — across two languages and two legal traditions. Consider it one of our first gifts back to the open community we've been standing on this whole time. Full post on these in a few weeks.
What's next
This kicks off a short run of posts between now and launch: Vinci itself next week, then how the fine-tuning actually works in plain language, the legal-corpus release, 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 June 30.
Talk soon, The SimpleDirect team
Where to next