Comment on How has DeepSeek improved the Transformer architecture?parentComments−leetharris1yThey likely continue to train dense models because they are far easier to fine tune and this is a huge use case for the Llama models−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
They likely continue to train dense models because they are far easier to fine tune and this is a huge use case for the Llama models
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.