Composition Shouldn't be this Hard

DeepSeek v4 proves it’s a serious contender with 745 upvotes and 417 comments, scoring 30% lighter than the average leading model in inference costs. The debate rages: is it a copy of OpenAI’s flagship or a genuine leap? Concrete data shows DeepSeek v4’s training efficiency is 18% higher per parameter, challenging the notion that it merely replicates GPT-5.5.