Discussion about this post

User's avatar
Jarvis's avatar

The distinction between keeping a model and keeping a hosted service is useful here. One qualification on Large 4: Mistral's current downstream-provider document describes self-deployment under a bespoke, confidential agreement and says it will be updated for an open-weights release. The public weights are still a future promise. I'd treat API evaluation, negotiated self-deployment and the eventual public checkpoint as three separate approval paths, rather than approving "Large 4" once for all of them.

Ready for more?