Really depends on the use case but one of the best all around models that’s multimodal, instruction tuned, has good tool support and json responses is gemma3-27b which you can run with about 48gb GPU mem.