Realistically you can assume there’ll be one soon and develop with these models and switch over
There are some already (one of the openllamas I think, and that discord group made one with the linear attention recently although trained on fewer params I think)