Comment on How has DeepSeek improved the Transformer architecture?parentComments−whimsicalism1yIt probably also has to do with their internal infra. If it were just about dense models being easier for the OSS community to use & build on, they should probably be training MoEs and then distilling to dense.
Comments
It probably also has to do with their internal infra. If it were just about dense models being easier for the OSS community to use & build on, they should probably be training MoEs and then distilling to dense.