DSpark: DeepSeek-V4's Insane Compute Optimization Explained
Check out Make and get a 1-month free Pro Plan with 10,000 operations & no credit card required! https://www.make.com/en?&promo=bycloud&utm_source=youtube&utm_medium=influencer&utm_campaign=bycloud-video-integration-jul26
They basically punched a hole in the ceiling of speculative decoding, and dropped DSpark. Never have I seen a lab sharing and even min-maxing inference this hard, and it is incredibly fascinating to see how they are approaching it not just from theory, but also from a practical serving perspective.
my latest project: Intuitive AI Academy
We just wrote a new piece on Opt