I could get PaliGemma working on RTX 3090 yesterday. Pretty good, probably the inference will be faster on RTX 4090. Maybe I will rent GPU and test that and other models today.