GritLM 7B is a text model that unifies embedding and generative capabilities in a single set of weights — a relatively rare combination. It can produce dense vector representations for retrieval tasks and generate text responses, switching between these modes without needing separate models. The trade-off is that it's a 7B parameter model, so raw generative power is modest compared to larger alternatives.