The release makes Model Runner V2 the default for two of the most widely deployed model families (Llama and Mistral), bringing its performance improvements — including pipeline-parallel bubble elimination and breakable CUDA graphs — to a much broader set of deployments.