This is a great strategy from Mistral AI: trying to own the open-source mixture-of-expert space.

MoE's are much easier to update with new abilities than having to finetune large models. Also faster and more composable.