Now that most foundation models are nearly satured into mature performance, I believe specialization ought to move forward from finetuning or just skill.md or design.md ... instead you could use Tetrees Agent where you can train and trade agentic intelligence packaged into .taip format. Tetrees Agent allows you to train intelligence with zero gpu, and be compatible with any llm including open weights. The training contains not just skills or toolsets but also experience of trial and error and failures ledger so you can skip them and expect better performance on each domains/subjects. The process you run afterward also gets reflected as solid experience and intelligence asset.