There are models with open weights, and you can run those locally on your GPU. It can be a bit slower depending on model and GPU. For example, GLM has an open version, both full and pruned, but it’s not the newest version. A bunch of image generation models have local versions too.