Provider Variation Surprise

Tom was surprised to discover up to 10% variation in benchmark outcomes for the same open-source model across different inference providers.