I don’t think it’s fair to think of model training as a one-and-done situation. It’s not like deepseek was designed and trained in one attempt. Every iteration of these models will require retraining until we have better continual learning implementations. Even when models are run locally, downloads signify demand, and demand calls for improved models, which means more training and testing is required.

That thing has an inbuilt DAC, as well as S/PDIF and USB inputs, so I’d imagine any device with enough processing power to play lossless audio files and has the proper drivers should be enough.