expectedwrong hindsight

Mistral Put a 7B Model on GitHub and Walked Away

A French startup just made the open-source licensing conversation significantly more awkward for Meta.

2 min read 267 words #open-source #llm #mistral #licensing #ml
hindsight — nailed it

Mistral is now valued at $14 billion with $400M ARR. The eight employees became 783. The "drop it and walk away" energy launched an entire company and an entire European AI industry. The Apache 2.0 license was the smartest business decision disguised as an act of generosity.

Mistral AI is about four months old. They have eight employees, give or take. They just posted a 7B model that beats Llama 2 13B on most benchmarks, under Apache 2.0, with a GitHub repo that is essentially: here is the model, here is the code, good luck.

No blog post. No waitlist. No usage policy document written by a committee. A torrent link and a README.

The Apache 2.0 part is what's going to stick. Meta shipped Llama 2 with a license that sounds open until you read it — commercial use gated on monthly active users, no using it to train other models, you're agreeing to a terms document that references Meta's acceptable use policy by hyperlink. Mistral looked at all of that and chose the license that has been saying yes to everything since 2004.

The benchmarks are almost beside the point, but they're also the point. A model that's 7 billion parameters beating one that's 13 billion isn't just a win for Mistral — it's a quiet suggestion that the race isn't over and the big guys don't have it figured out. Sliding window attention, grouped-query attention, some implementation choices that apparently matter a lot. The details are in the repo. The repo is right there.

There's a version of the next two years where this is a footnote. There's another version where September 27th, 2023 is a date people remember.

I'm not making a prediction. I'm just noting that a very small group of people in Paris apparently decided that the correct move was to ship the thing, and they were right.