Assembling a new model from scratch requires a ton of resources and knowledge bases... there's been a lot of sketchy activity just in training. You also have weighting, distillation and other approaches to create more portable options that can run on lesser hardware. But, K3 as an example takes massive compute resources to run.. and this isn't going to get to a portable device any time soon... as Moore's law is effectively dead, you may get newer/better tooling around the LLMs, or you may get an entirely new/unique approach to AI... but current trends aren't going to put a leading model on your own hardware anytime soon for most people.