Theory of the evolution of pathogen specialization suggests that a specialist pathogen gains high fitness in one host, but this comes with fitness loss in other hosts. By contrast, a generalist pathogen does not achieve high fitness in any host, but gains ecological fitness by exploiting different hosts, and has higher fitness than specialists in nonspecialized hosts. As a result, specialist pathogens are predicted to have greater variation in fitness across hosts, and generalists would have lower fitness variation across hosts. We test these hypotheses by measuring pathogen replicative fitness as within-host viral loads from the onset of infection to the beginning of virus clearance, using the rhabdovirus infectious hematopoietic necrosis virus (IHNV) in salmonid fish. Based on field prevalence and virulence studies, the IHNV subgroups UP, MD, and L are specialists, causing infection and mortality in sockeye salmon, steelhead, and Chinook salmon juveniles, respectively. The UC subgroup evolved naturally from a UP ancestor and is a generalist infecting all three host species but without causing severe disease. We show that the specialist subgroups had the highest peak and mean viral loads in the hosts in which they are specialized, and they had low viral loads in nonspecialized hosts, resulting in large variation in viral load across hosts. Viral kinetics show that the mechanisms of specialization involve the ability to both maximize early virus replication and avoid clearance at later times, with different mechanisms of specialization evident in different host–virus combinations. Additional nuances in the data included different fitness levels for nonspecialist interactions, reflecting different trade-offs for specialist viruses in other hosts. The generalist UC subgroup reached intermediate viral loads in all hosts and showed the smallest variation in fitness across hosts. The evolution of the UC generalist from an ancestral UP sockeye specialist was associated with fitness increases in steelhead and Chinook salmon, but only slight decreases in fitness in sockeye salmon, consistent with low- or no-cost generalism. Our results support major elements of the specialist–generalist theory, providing evidence of a specialist–generalist continuum in a vertebrate pathogen. These results also quantify within-host replicative fitness trade-offs resulting from the natural evolution of specialist and generalist virus lineages in multi-host ecosystems