They also publish papers talking about how to save kv cache and computation powers. Because currently they don't have the most powerful nvidia cards, training and inference efficiency is very import for them.

1) someone deepseeks deepseek lol:

Generates their own weights and figures out a way to determine all of the intermediate states.

2) places realize there’s real risk with using a model that might have things baked into it that produce specific flaws that could be security bugs, but only under certain conditions.