I'd say for research purposes. HackerNews seems to tend to mostly represent the LLM consumer viewpoint, but these waves of models being released are honestly more interesting from a research than a user point of view. As a LLM user, you're (vast generalization/simplification) really just interested in the best of N models, but as a researcher, I'm super interested in each model's performance and analyzing the reasons for any differences.

With this model (and they say this in the blog post), they were testing the hypothesis that training on a longer context size would provide more performance at the same parameter count/inference FLOPs. From a quick perusal of their post, it looks like this was true, and we should train all future models with as long of a context size as we can afford.