Mila splits the open AI foundation layer into an interface standard and a reference install

Whether you can pull a piece out of this stack and drop in a better one comes down to the published interface contracts. Mozilla says the initiative has two halves. One is an open standard published as interface contracts, so any part of the stack can be swapped for a better one. The other is a working reference implementation an organization can install on its own machines, running the models and controls it chooses against its own data, with governance and access control built in from the start ^[9]. Contracts with versions are what let you replace an inference server or a vector store and keep everything above it. The announcement is silent on the repository, the license, the spec version and the delivery date ^[19].

By Mozilla's own account, the reference implementation is the expensive half. Using open source models was never the hard part. Turning a raw model into something a small business, a hospital, a local charity or a government can run in production takes an engineering team most organizations do not have and months they cannot spare ^[10]. The worked example in the post is a small manufacturer running a private assistant over its manuals, procedures and past project files. It helps staff draft reports and answer technical questions, and helps its software team write code ^[11]. That means a retrieval index over messy internal documents plus per-user permissions on the files it returns.

Mozilla provided the initial $5 million and contributes technical expertise, while Mila leads technical delivery and coordination ^[4][3]. Hypertec is committing a further $1 million in first-year funding to accelerate initial Canadian deployments of the layer on Hypertec hardware ^[5]. The public commitments total $6 million ^[17], and one sixth of that is contingent on deployments landing on one supplier's machines ^[18]. For the portability the standard promises to transfer, the reference install has to be exercised on other suppliers' hardware.

"Canada has a choice: depend on technologies developed elsewhere, or build more of what we need here at home," said Evan Solomon, Minister of Artificial Intelligence and Digital Innovation ^[6][7]. The Government of Canada announced its support for the initiative ^[2]. Valerie Pisano, President and CEO of Mila, said: "Six months ago, we announced our partnership with Mozilla to advance open source and sovereign AI. Today, we are bringing that work to a whole new level." ^[13] Mila brings close to 2,000 researchers and professionals; Mozilla brings 25 years stewarding open infrastructure others build on ^[8].

Mozilla writes that for the vast majority of everyday business tasks, companies will be able to use open source AI running on open source tools instead of paying a commercial provider every time an employee makes a request. That depends, Mozilla writes, on reducing the technical work and expense involved in putting open source AI into practice ^[12]. A per-request comparison only favours ownership once you also count the GPUs, the person who patches them, and the hours the hardware sits idle. Rented capacity bills only when it is used. Mozilla states the goal as making owning your AI as easy as renting it ^[15], and the two organizations are inviting companies, research institutions, funders, governments and developers to join the work ^[16].