Mistral Large 4 uses a sparse mixture-of-experts architecture that activates approximately 50 billion parameters per token