Essay · Open weights
The case for sovereign weights, written from a small office in Paris
Why a continent should own at least one model end-to-end — and what we learned shipping our second mixture-of-experts release on a March deadline we very nearly missed.
There is a quiet office on rue Saint-Honoré where, last winter, four of us spent two weeks arguing about whether the dataset filter should drop a French-language corpus that nobody outside the lab had ever asked us about. We kept it. The corpus was small, idiosyncratic, full of legal-aid pamphlets and post-1995 administrative French — exactly the kind of thing a continental model should be embarrassed to lose. The eval needle barely moved. We shipped anyway.
A weight that lives on your own machine is a different kind of commitment
Sovereignty is a word that gets carried around like a flag, but the engineering version of it is duller and more interesting. It is: who is allowed to read the gradient logs at 3am, whose subpoena reaches the storage bucket, and which jurisdiction owns the silicon that ran the last training step. Renting a model behind an opaque endpoint answers none of those. Publishing the weights answers all of them at once, and creates four new problems by Tuesday.
We released the 8×7B checkpoint in December under a permissive license, and within a week a hospital in Lyon had it running on two consumer GPUs to triage radiology notes — no API, no audit logs leaving the building. That is the entire argument, condensed.