What We're Building: An Open-Weight Canadian Model Series

The model is the smallest part of the story. Here's what it is, what it isn't, and what comes next.

Today we shipped flash-1-mini.

It's a 4-billion-parameter open-weight model, fine-tuned for Canadian context, bilingual in English and French, that runs on a laptop with no cloud dependency.

You can download it, run it offline, and own it.

The weights are yours.

I want to write about what it is, what it isn't, and what comes after — because the model itself is the smallest part of the story.

The empty slot in the map

Go look at the landscape of AI models built for government and regulated workflows. Here's what you find.

In the United States, defense AI is dominated by two players.

Scale AI's Defense Llama — a Llama 3 fine-tune that runs inside Scale's Donovan platform on classified networks.

And IBM's defense model — built on Granite, deployable in air-gapped environments, sold by subscription.

Both are proprietary.

Both are locked behind classification or a paywall.

Neither is publicly accessible.

In Canada, the federal government's CANChat platform runs a mix of foreign-trained foundation models with retrieval over Canadian government documents.

The public framing positions it as sovereign Canadian AI.

The operational reality, surfaced through public records, is more nuanced.

The foundation models powering it weren't trained on Canadian data.

They were trained by Meta, Cohere, OpenAI, Google.

Canada is the wrapper, not the foundation.

Cohere, Canada's domestic frontier company, runs closed commercial models.

You can use them through their API. You can't download them, run them locally, or audit them.

The weights aren't yours.

That leaves a gap that's actually quite precise.

There is no open-weight, publicly accessible, Canadian-context-trained model that anyone can download, run on their own hardware, and verify.

Not in defense. Not in government. Not in the public domain.

flash-1-mini is the first step toward filling it.

What flash-1-mini is

A 4-billion-parameter fine-tune of a strong open-weight base model on the frontier of intelligence-per-parameter, trained on Canadian regulatory, legal, and bilingual context.

About 2.7 GB on disk.

It runs on a MacBook Air with 16GB of unified memory.

It runs on a Raspberry Pi 5, with some patience.

What does it do better than the base model?

On our internal benchmarks: 10–15 points of improvement across Canadian legal reasoning, bilingual fluency (especially Quebec French legal terminology), tool calling for agentic workflows, and faithfulness when retrieving over Canadian regulatory text.

It's not magic. It's a base model that's been taught to think about Canada the way the base model wasn't.

It will not replace Claude or GPT for general-purpose chat. It isn't trying to.

It's the model you reach for when the context is Canadian law, Canadian regulation, Canadian bilingual administrative work — or any deployment where running locally, with no data leaving your hardware, is a feature and not a constraint.

What flash-1-mini isn't

It is not a finished product. It's the first member of a family.

It is not a competitor to closed frontier models on general intelligence.

It can't match the mathematical reasoning of GPT or Claude.

It's one-twentieth their parameter count, by design.

It is not the most accurate model for English-language Canadian context.

The flash-1 release (9B, July) will be measurably better.

flash-1-pro (27B, September) will be better still.

It is not yet specialized for any single high-stakes domain.

The model tuned specifically for Government of Canada workflows — the one I'm most excited about — comes after September, once the base capability is locked.

It's a starting point.

The receipt that says we can do this, and here's what it looks like.

The warmup.

The roadmap

We're going to ship four open-weight models in 2026, then one specialized variant.

Same deal every time: you buy it, you own it, you run it on hardware you control. No subscription. No per-token meter.

The open dataset — alongside flash-1 (July 2026) — free

The bilingual Canadian fine-tuning dataset we used, published openly.

Train against it. Benchmark against it.

We want this to become the standard for how Canadian-context AI gets evaluated — because that standard should exist and currently doesn't.

The government variant — after September 2026 — free public demo

A fine-tune built on government documents, hosted so any public servant, policy analyst, or curious citizen can try it in the browser.

Not a product.

A demonstration of what a Canadian-built AI answering questions about Canadian government — in both official languages, locally, without phoning home to Silicon Valley — actually looks like.

Why we're building it this way

Let me be specific about the philosophy, because the model is the smallest part.

The closed-source labs aren't wrong about what they're building.

OpenAI and Anthropic are pushing the frontier on raw capability, pulling in tens of billions of dollars and pointing it at training runs nobody else can afford.

If you want the smartest possible model, you call them. That's correct.

But there's a separate question about who owns the AI you depend on.

And that question is being answered by default, every day, by every team that wires its workflow to an API someone else controls — in a jurisdiction that isn't ours, with weights they can't see and training data they can't audit.

When the Government of Canada deploys AI to help public servants, the model running underneath is, operationally, today, a foreign foundation model wrapped in Canadian retrieval.

That isn't sovereign AI. That's depending on someone else's sovereign AI with a Canadian layer on top.

We think there's room for a different shape.

Open-weight, locally deployable, Canadian-trained models that the people using them can verify, audit, modify, and run on their own hardware, in their own country.

That's what an "AI you can own" series looks like.

flash-1-mini is the first one.

Who this is for

Let me be honest about who this is and isn't for.

It's not for you if you need GPT-grade general reasoning. We're not in that fight.

It's not for you if you're happy with cloud AI subscriptions and the model running on someone else's hardware is fine.

There's a real efficiency argument for that, and we're not going to talk you out of it.

It is for you if you're a Canadian small business that wants AI in your workflow without depending on a US API that could change pricing, terms, or availability at any time.

It is for you if you're a policy analyst, lawyer, or public servant who wants to feel what locally running bilingual AI is like before recommending it to a department.

It is for you if you're building edge deployments — local inference from Raspberry Pi-class hardware up to small servers — where cloud APIs are operationally impossible.

And it is for you if "what do you actually own in the AI era" is a question you've been asking yourself, and you want to see what an answer looks like.

What's next

flash-1-mini is up now.

Hugging Face for the weights.

GitHub for the methodology, model card, and benchmarks.

Founder Reality and SimpleDirect for the thinking behind the work.

The next member lands in July, anchored by the dataset release.

Then flash-1-pro in September. Then the variant.

We're a small team team building this in Toronto.

Not VC-funded. Not a SaaS platform. Not a portal.

We ship models, release datasets, and publish methodology — because the world needs more open-weight Canadian-context AI in it, and nobody else was making it.

You don't have to take our word for any of this.

The whole point of open weights is that you can download, inspect, and run.

So do that. Then tell us what we got wrong.

We'll fix it in the next release.

— George