Trending: On-device modelsSearch
iHeartGeek
iTECH

Mila and Mozilla build a sovereign open source AI stack

Mozilla's first partnership with a major AI research lab puts $5m behind a foundation layer that organisations can run on their own hardware, with Canada's government on board.

Abstract editorial illustration of a modular stack of translucent interchangeable layers, linked by fine circuit traces and rising from a solid base block.

Mila and Mozilla have announced a joint initiative to build an open source AI foundation layer that an organisation can install and run on its own machines. The announcement was made at ALL IN, Canada's largest AI and technology event. Mozilla is putting in the initial $5 million and acting as technical partner; Mila, with close to 2,000 researchers and professionals behind it, leads the technical delivery and coordination. The Government of Canada has backed the work, and Hypertec is adding $1 million of first-year funding to get early Canadian deployments running on its hardware.

What a foundation layer actually means

It has two halves. The first is an open standard, published as interface contracts, so that any part of the stack can be swapped for something better without rebuilding the rest. The second is a working reference implementation: a package any organisation can install on its own machines, running the models, tools and access controls it chooses, against its own data, with governance built in from the start. The stated goal is to make owning your AI roughly as easy as renting it.

The gap they are aiming at

Mila and Mozilla point at a familiar problem. Plenty of teams now work with open models, but getting one into production — secure, reliable and connected to everything else — is a project most organisations cannot staff. Their example is a small manufacturer running a private assistant that searches its own manuals and project files, drafts reports and helps its software team write code, without any of that leaving the building. A hospital or a charity could run the same thing, and the research partnership's opening project will look at trust and usability, including private memory architectures for agents. The partnership is also Mozilla's first with a major AI research lab.

Where the money and the politics land

"Canada has a choice: depend on technologies developed elsewhere, or build more of what we need here at home," said Evan Solomon, the country's Minister of Artificial Intelligence and Digital Innovation. Valérie Pisano, Mila's president and chief executive, framed the work as letting organisations "run, control, and maintain AI models themselves", and said the pair have spent six months designing the architecture and testing it end to end. Hypertec's chief executive, Simon Ahdoot, gets the commercial end: his company supplies the hardware the first Canadian deployments will run on.

Our opinion

"Sovereign AI" has become the phrase every government and enterprise vendor reaches for in 2026, and most of what is sold under it is a licensing arrangement with better paperwork. The parts of this one that are worth watching are the interface contracts and the reference implementation, because those are the pieces that decide whether an organisation with two engineers can actually run the thing. Mozilla's twenty-five years of keeping open infrastructure alive is the right kind of credibility for a standards play, and Mila's research community is a real one. The test is adoption: if the standard ends up with contributors outside Canada and implementations nobody had to ask permission for, it will have mattered. If it stays a Canadian government programme with a Mozilla logo on it, it will be a press release with a longer tail.

What we know
  • Mozilla funds the first $5m; Mila leads technical delivery
  • Open standard plus a reference implementation organisations install themselves
  • Government of Canada backing; Hypertec adds $1m for early Canadian deployments